MVAPICH2 1.9 User Guide

MVAPICH Team
Network-Based Computing Laboratory
Department of Computer Science and Engineering
The Ohio State University
http://mvapich.cse.ohio-state.edu
Copyright (c) 2001-2014
Network-Based Computing Laboratory,
headed by Dr. D. K. Panda.
All rights reserved.

Last revised: January 3, 2014

Contents

1 Overview of the MVAPICH Project
2 How to use this User Guide?
3 MVAPICH2 1.9 Features
4 Installation Instructions
 4.1 Building from a tarball
 4.2 Obtaining and Building the Source from SVN repository
 4.3 Selecting a Process Manager
  4.3.1 Customizing Commands Used by mpirun_rsh
  4.3.2 Using SLURM
 4.4 Configuring a build for OFA-IB-CH3/OFA-iWARP-CH3/OFA-RoCE-CH3
 4.5 Configuring a build for NVIDIA GPU with OFA-IB-CH3
 4.6 Configuring a build for Shared-Memory-CH3
 4.7 Configuring a build for OFA-IB-Nemesis
 4.8 Configuring a build for uDAPL-CH3
 4.9 Configuring a build for QLogic PSM-CH3
 4.10 Configuring a build for TCP/IP-Nemesis
 4.11 Configuring a build for TCP/IP-CH3
 4.12 Configuring a build for OFA-IB-Nemesis and TCP/IP Nemesis (unified binary)
 4.13 Configuring a build for Shared-Memory-Nemesis
5 Basic Usage Instructions
 5.1 Compile Applications
 5.2 Run Applications
  5.2.1 Run using mpirun_rsh
  5.2.2 Run using Hydra (mpiexec)
  5.2.3 Run using SLURM
  5.2.4 Run on PBS/Torque Clusters
  5.2.5 Run with Dynamic Process Management support
  5.2.6 Run with mpirun_rsh using OFA-iWARP Interface
  5.2.7 Run with mpirun_rsh using OFA-RoCE Interface
  5.2.8 Run using Hydra (mpiexec) with uDAPL-CH3 Interface
  5.2.9 Run using IPoIB with mpirun_rsh or mpiexec
  5.2.10 Run using ADIO driver for Lustre
  5.2.11 Run using TotalView Debugger Support
  5.2.12 Run using a profiling library
6 Advanced Usage Instructions
 6.1 Running on Customized Environments
 6.2 Export Environment
  6.2.1 Sample Use
 6.3 Configuration File Processing
  6.3.1 Sample Use
 6.4 Suspend/Resume Support
 6.5 Running with Efficient CPU (Core) Mapping
  6.5.1 Using HWLOC for CPU Mapping
  6.5.2 User defined CPU Mapping
  6.5.3 Performance Impact of CPU Mapping
 6.6 Running with LiMIC2
 6.7 Running with Shared Memory Based Windows for One-sided Communication
 6.8 Running with Shared Memory Collectives
 6.9 Running Collectives with Hardware based Multicast support
 6.10 Running MPI_Gather collective with intra-node Zero-Copy designs (using LiMIC2)
 6.11 Running with scalable UD transport
 6.12 Running with Integrated Hybrid UD-RC/XRC design
 6.13 Running with Multiple-Rail Configurations
 6.14 Enhanced design for Multiple-Rail Configurations
 6.15 Running with Fault-Tolerance Support
  6.15.1 System-Level Checkpoint/Restart
  6.15.2 Multi-Level Checkpointing with Scalable Checkpoint-Restart (SCR)
  6.15.3 Job Pause-Migration-Restart Support
  6.15.4 Run-Through Stabilization
  6.15.5 Network Fault Tolerance with Automatic Path Migration
 6.16 Running with RDMA CM support
 6.17 Running MVAPICH2 in Multi-threaded Environments
 6.18 Compiler Specific Flags to enable OpenMP thread binding
 6.19 Running with Hot-Spot and Congestion Avoidance
 6.20 Running on Clusters with NVIDIA GPU Accelerators
 6.21 MPIRUN_RSH compatibility with MPIEXEC
  6.21.1 Interaction with SLURM
  6.21.2 Interaction with PBS
7 OSU Benchmarks
 7.1 Download and Build Stand-alone OSU Benchmarks Package
 7.2 Running
  7.2.1 Running OSU Latency and Bandwidth
  7.2.2 Running OSU Message Rate Benchmark
  7.2.3 Running OSU Collective Benchmarks
  7.2.4 Running Benchmarks with CUDA/OpenACC Extensions
8 Scalability features and Performance Tuning for Large Scale Clusters
 8.1 Optimizations for homogeneous clusters
 8.2 Job Launch Tuning
 8.3 Basic QP Resource Tuning
 8.4 RDMA Based Point-to-Point Tuning
 8.5 Shared Receive Queue (SRQ) Tuning
 8.6 eXtended Reliable Connection (XRC)
 8.7 Shared Memory Tuning
 8.8 On-demand Connection Management Tuning
 8.9 Scalable Collectives Tuning
  8.9.1 Optimizations for MPI_Bcast
  8.9.2 Optimizations for MPI_Reduce and MPI_Allreduce
  8.9.3 Optimizations for MPI_Gather and MPI_Scatter
 8.10 Process Placement on Multi-core platforms
 8.11 HugePage Support
9 FAQ and Troubleshooting with MVAPICH2
 9.1 General Questions and Troubleshooting
  9.1.1 MVAPICH2 is failed to register memory with Infiniband HCA
  9.1.2 Invalid Communicators Error
  9.1.3 Are fork() and system() supported?
  9.1.4 MPI+OpenMP shows bad performance
  9.1.5 Error message “No such file or directory” when using Lustre file system
  9.1.6 Program segfaults with “File locking failed in ADIOI_Set_lock”
  9.1.7 Running MPI programs built with gfortran
  9.1.8 How do I obtain MVAPICH2 version and configuration information?
  9.1.9 How do I compile my MPI application with static libraries, and not use shared libraries?
  9.1.10 Does MVAPICH2 work across AMD and Intel systems?
  9.1.11 I want to enable debugging for my build. How do I do this?
  9.1.12 How can I run my application with a different group ID?
 9.2 Issues and Failures with Job launchers
  9.2.1 Can I use MVAPICH2 on PBS/Torque Clusters
  9.2.2 /usr/bin/env: mpispawn: No such file or directory
  9.2.3 TotalView complains that “The MPI library contains no suitable type definition for struct MPIR_PROCDESC”
 9.3 Problems Building MVAPICH2
  9.3.1 Unable to convert MPI_SIZEOF_AINT to a hex string
  9.3.2 Cannot Build with the PathScale Compiler
  9.3.3 Libtool has a problem linking with non-GNU compiler (like PGI)
 9.4 With OFA-IB-CH3 Interface
  9.4.1 Cannot Open HCA
  9.4.2 Checking state of IB Link
  9.4.3 Creation of CQ or QP failure
  9.4.4 Hang with the HSAM Functionality
  9.4.5 Failure with Automatic Path Migration
  9.4.6 Error opening file
  9.4.7 RDMA CM Address error
  9.4.8 RDMA CM Route error
 9.5 With OFA-iWARP-CH3 Interface
  9.5.1 Error opening file
  9.5.2 RDMA CM Address error
  9.5.3 RDMA CM Route error
 9.6 With uDAPL-CH3 Interface
  9.6.1 Cannot Open IA
  9.6.2 DAT Insufficient Resource
  9.6.3 Cannot Find libdat.so
  9.6.4 uDAPL over IB Does Not Scale Beyond 256 Nodes with RDMA_CM Provider
 9.7 Checkpoint/Restart
  9.7.1 Failure during Restart
10 MVAPICH2 General Parameters
 10.1 MV2_IGNORE_SYSTEM_CONFIG
 10.2 MV2_IGNORE_USER_CONFIG
 10.3 MV2_USER_CONFIG
 10.4 MV2_DEBUG_CORESIZE
 10.5 MV2_DEBUG_SHOW_BACKTRACE
11 MVAPICH2 Parameters (CH3-Based Interfaces)
 11.1 MV2_ALLREDUCE_2LEVEL_MSG
 11.2 MV2_CKPT_AGGREGATION_BUFPOOL_SIZE
 11.3 MV2_CKPT_AGGREGATION_CHUNK_SIZE
 11.4 MV2_CKPT_FILE
 11.5 MV2_CKPT_INTERVAL
 11.6 MV2_CKPT_MAX_SAVE_CKPTS
 11.7 MV2_CKPT_NO_SYNC
 11.8 MV2_CKPT_USE_AGGREGATION
 11.9 MV2_DEBUG_FT_VERBOSE
 11.10 MV2_CM_RECV_BUFFERS
 11.11 MV2_CM_SPIN_COUNT
 11.12 MV2_CM_TIMEOUT
 11.13 MV2_CPU_MAPPING
 11.14 MV2_CPU_BINDING_POLICY
 11.15 MV2_CPU_BINDING_LEVEL
 11.16 MV2_SHOW_CPU_BINDING
 11.17 MV2_DAPL_PROVIDER
 11.18 MV2_DEFAULT_MAX_SEND_WQE
 11.19 MV2_DEFAULT_MAX_RECV_WQE
 11.20 MV2_DEFAULT_MTU
 11.21 MV2_DEFAULT_PKEY
 11.22 MV2_ENABLE_AFFINITY
 11.23 MV2_GET_FALLBACK_THRESHOLD
 11.24 MV2_IBA_EAGER_THRESHOLD
 11.25 MV2_IBA_HCA
 11.26 MV2_INITIAL_PREPOST_DEPTH
 11.27 MV2_IWARP_MULTIPLE_CQ_THRESHOLD
 11.28 MV2_KNOMIAL_INTRA_NODE_FACTOR
 11.29 MV2_KNOMIAL_INTER_NODE_FACTOR
 11.30 MV2_LIMIC_GET_THRESHOLD
 11.31 MV2_LIMIC_PUT_THRESHOLD
 11.32 MV2_MAX_INLINE_SIZE
 11.33 MV2_MAX_NUM_WIN
 11.34 MV2_NDREG_ENTRIES
 11.35 MV2_NUM_HCAS
 11.36 MV2_NUM_PORTS
 11.37 MV2_NUM_SA_QUERY_RETRIES
 11.38 MV2_NUM_QP_PER_PORT
 11.39 MV2_RAIL_SHARING_POLICY
 11.40 MV2_RAIL_SHARING_LARGE_MSG_THRESHOLD
 11.41 MV2_PROCESS_TO_RAIL_MAPPING
 11.42 MV2_RDMA_FAST_PATH_BUF_SIZE
 11.43 MV2_NUM_RDMA_BUFFER
 11.44 MV2_ON_DEMAND_THRESHOLD
 11.45 MV2_HOMOGENEOUS_CLUSTER
 11.46 MV2_PREPOST_DEPTH
 11.47 MV2_PREPOST_DEPTH
 11.48 MV2_PROCESS_TO_RAIL_MAPPING
 11.49 MV2_PSM_DEBUG
 11.50 MV2_PSM_DUMP_FREQUENCY
 11.51 MV2_PUT_FALLBACK_THRESHOLD
 11.52 MV2_RAIL_SHARING_LARGE_MSG_THRESHOLD
 11.53 MV2_RAIL_SHARING_POLICY
 11.54 MV2_RDMA_CM_ARP_TIMEOUT
 11.55 MV2_RDMA_CM_MAX_PORT
 11.56 MV2_RDMA_CM_MIN_PORT
 11.57 MV2_REDUCE_2LEVEL_MSG
 11.58 MV2_RNDV_PROTOCOL
 11.59 MV2_R3_THRESHOLD
 11.60 MV2_R3_NOCACHE_THRESHOLD
 11.61 MV2_SHMEM_ALLREDUCE_MSG
 11.62 MV2_SHMEM_BCAST_LEADERS
 11.63 MV2_SHMEM_BCAST_MSG
 11.64 MV2_SHMEM_COLL_MAX_MSG_SIZE
 11.65 MV2_SHMEM_COLL_NUM_COMM
 11.66 MV2_SHMEM_DIR
 11.67 MV2_SHMEM_REDUCE_MSG
 11.68 MV2_SM_SCHEDULING
 11.69 MV2_SMP_USE_LIMIC2
 11.70 MV2_SMP_USE_CMA
 11.71 MV2_SRQ_LIMIT
 11.72 MV2_SRQ_MAX_SIZE
 11.73 MV2_SRQ_SIZE
 11.74 MV2_STRIPING_THRESHOLD
 11.75 MV2_SUPPORT_DPM
 11.76 MV2_USE_APM
 11.77 MV2_USE_APM_TEST
 11.78 MV2_USE_BLOCKING
 11.79 MV2_USE_COALESCE
 11.80 MV2_USE_DIRECT_GATHER
 11.81 MV2_USE_DIRECT_SCATTER
 11.82 MV2_USE_HSAM
 11.83 MV2_USE_IWARP_MODE
 11.84 MV2_USE_LAZY_MEM_UNREGISTER
 11.85 MV2_USE_LIMIC_ONE_SIDED
 11.86 MV2_USE_RoCE
 11.87 MV2_DEFAULT_GID_INDEX
 11.88 MV2_USE_RDMA_CM
 11.89 MV2_RDMA_CM_CONF_FILE_PATH
 11.90 MV2_USE_RDMA_FAST_PATH
 11.91 MV2_USE_RDMA_ONE_SIDED
 11.92 MV2_USE_RING_STARTUP
 11.93 MV2_USE_SHARED_MEM
 11.94 MV2_USE_SHM_ONE_SIDED
 11.95 MV2_USE_SHMEM_ALLREDUCE
 11.96 MV2_USE_SHMEM_BARRIER
 11.97 MV2_USE_SHMEM_BCAST
 11.98 MV2_USE_SHMEM_COLL
 11.99 MV2_USE_SHMEM_REDUCE
 11.100 MV2_USE_SRQ
 11.101 MV2_GATHER_SWITCH_PT
 11.102 MV2_SCATTER_SMALL_MSG
 11.103 MV2_SCATTER_MEDIUM_MSG
 11.104 MV2_USE_TWO_LEVEL_GATHER
 11.105 MV2_USE_TWO_LEVEL_SCATTER
 11.106 MV2_USE_XRC
 11.107 MV2_VBUF_POOL_SIZE
 11.108 MV2_VBUF_SECONDARY_POOL_SIZE
 11.109 MV2_VBUF_TOTAL_SIZE
 11.110 MV2_SMP_EAGERSIZE
 11.111 MV2_SMPI_LENGTH_QUEUE
 11.112 MV2_SMP_NUM_SEND_BUFFER
 11.113 MV2_SMP_SEND_BUF_SIZE
 11.114 MV2_USE_HUGEPAGES
 11.115 MV2_HYBRID_ENABLE_THRESHOLD
 11.116 MV2_HYBRID_MAX_RC_CONN
 11.117 MV2_UD_PROGRESS_TIMEOUT
 11.118 MV2_UD_RETRY_TIMEOUT
 11.119 MV2_UD_RETRY_COUNT
 11.120 MV2_USE_UD_HYBRID
 11.121 MV2_USE_ONLY_UD
 11.122 MV2_USE_UD_ZCOPY
 11.123 MV2_USE_LIMIC_GATHER
 11.124 MV2_USE_MCAST
 11.125 MV2_MCAST_NUM_NODES_THRESHOLD
 11.126 MV2_USE_CUDA
 11.127 MV2_CUDA_BLOCK_SIZE
 11.128 MV2_CUDA_IPC
 11.129 MV2_CUDA_SMP_IPC
12 MVAPICH2 Parameters (OFA-IB-Nemesis Interface)
 12.1 MV2_DEFAULT_MAX_SEND_WQE
 12.2 MV2_DEFAULT_MAX_RECV_WQE
 12.3 MV2_DEFAULT_MTU
 12.4 MV2_DEFAULT_PKEY
 12.5 MV2_IBA_EAGER_THRESHOLD
 12.6 MV2_IBA_HCA
 12.7 MV2_INITIAL_PREPOST_DEPTH
 12.8 MV2_MAX_INLINE_SIZE
 12.9 MV2_NDREG_ENTRIES
 12.10 MV2_NUM_RDMA_BUFFER
 12.11 MV2_NUM_SA_QUERY_RETRIES
 12.12 MV2_PREPOST_DEPTH
 12.13 MV2_RNDV_PROTOCOL
 12.14 MV2_R3_THRESHOLD
 12.15 MV2_R3_NOCACHE_THRESHOLD
 12.16 MV2_SRQ_LIMIT
 12.17 MV2_SRQ_SIZE
 12.18 MV2_STRIPING_THRESHOLD
 12.19 MV2_USE_BLOCKING
 12.20 MV2_USE_LAZY_MEM_UNREGISTER
 12.21 MV2_USE_RDMA_FAST_PATH
 12.22 MV2_USE_SRQ
 12.23 MV2_VBUF_POOL_SIZE
 12.24 MV2_VBUF_SECONDARY_POOL_SIZE
 12.25 MV2_VBUF_TOTAL_SIZE
 12.26 MV2_RUN_THROUGH_STABILIZATION
13 MPIRUN_RSH Parameters
 13.1 MV2_COMM_WORLD_LOCAL_RANK
 13.2 MV2_COMM_WORLD_LOCAL_SIZE
 13.3 MV2_COMM_WORLD_RANK
 13.4 MV2_COMM_WORLD_SIZE
 13.5 MV2_FASTSSH_THRESHOLD
 13.6 MV2_NPROCS_THRESHOLD
 13.7 MV2_MPIRUN_TIMEOUT
 13.8 MV2_MT_DEGREE
 13.9 MPIEXEC_TIMEOUT
 13.10 MV2_DEBUG_FORK_VERBOSE

1 Overview of the MVAPICH Project

InfiniBand, 10GigE/iWARP and 10/40GigE RDMA over Converged Ethernet (RoCE) are emerging as high-performance networking technologies to deliver low latency and high bandwidth. They are also achieving widespread acceptance due to their open standards.

MVAPICH (pronounced as “em-vah-pich”) is an open-source MPI software to exploit the novel features and mechanisms of these networking technologies and deliver best performance and scalability to MPI applications. This software is developed in the Network-Based Computing Laboratory (NBCL), headed by Prof. Dhabaleswar K. (DK) Panda.

The MVAPICH2 MPI library supports MPI-3 semantics. This open-source MPI software project started in 2001 and a first high-performance implementation was demonstrated at Supercomputing ’02 conference. After that, this software has been steadily gaining acceptance in the HPC, InfiniBand, 10GigE/iWARP and 10/40GigE RoCE communities. As of May 6, 2013, more than 2,055 organizations (National Labs, Universities and Industry) world-wide (in 70 countries) have registered as MVAPICH users at MVAPICH project web site. There have also been more than 167,000 downloads of this software from the MVAPICH project site directly. In addition, many InfiniBand, 10GigE/iWARP and 10/40GigE RoCE vendors, server vendors, systems integrators and Linux distributors have been incorporating MVAPICH2 into their software stacks and distributing it. MVAPICH2 distribution is available under BSD licensing.

Several InfiniBand systems using MVAPICH2 have obtained positions in the TOP 500 ranking. The November ’12 list includes the following systems: 7th ranked Stampede system at TACC with 204,900-cores, 14th ranked Pleiades system at NASA with 125,980 cores, 17th ranked Tsubame2 system at Tokyo Institute of Technology with 73,278 cores, and 50th ranked Ranger system at Texas Advanced Computing Center (TACC) with 62,976 cores.

More details on MVAPICH software, users list, mailing lists, sample performance numbers on a wide range of platforms and interconnects, a set of OSU benchmarks, related publications, and other InfiniBand- and iWARP-related projects (parallel file systems, storage, data centers) can be obtained from our website:http://mvapich.cse.ohio-state.edu.

This document contains necessary information for MVAPICH2 users to download, install, test, use, tune and troubleshoot MVAPICH2 1.9. We continuously fix bugs and update update this document as per user feedback. Therefore, we strongly encourage you to refer to our web page for updates.

2 How to use this User Guide?

This guide is designed to take the user through all the steps involved in configuring, installing, running and tuning MPI applications over InfiniBand using MVAPICH2 1.9.

In Section 3 we describe all the features in MVAPICH2 1.9. As you read through this section, please note our new features (highlighted as NEW) compared to version 1.8. Some of these features are designed in order to optimize specific type of MPI applications and achieve greater scalability. Section 4 describes in detail the configuration and installation steps. This section enables the user to identify specific compilation flags which can be used to turn some of the features on of off. Basic usage of MVAPICH2 is explained in Section 5. Section 6 provides instructions for running MVAPICH2 with some of the advanced features. Section 7 describes the usage of the OSU Benchmarks. In Section 8 we suggest some tuning techniques for multi-thousand node clusters using some of our new features. If you have any problems using MVAPICH2, please check Section 9 where we list some of the common problems people face. Finally, in Sections 11 and 12, we list all important run time parameters, their default values and a short description.

3 MVAPICH2 1.9 Features

MVAPICH2 (MPI-3 over InfiniBand) is an MPI-3 implementation based on MPICH ADI3 layer. MVAPICH2 1.9 is available as a single integrated package (with MPICH 3.0.3). The current release supports ten different underlying transport interfaces, as shown in Figure 1.


PIC

Figure 1: Overview of different available interfaces of the MVAPICH2 library


MVAPICH2 1.9 is compliant with MPI 3 standard. In addition, MVAPICH2 1.9 provides support and optimizations for NVIDIA GPU, multi-threading and fault-tolerance (Checkpoint-restart, Job-pause-migration-resume). A complete set of features of MVAPICH2 1.9 are indicated below. New features compared to 1.8 are indicated as (NEW).

The MVAPICH2 1.9 package and the project also includes the following provisions:

4 Installation Instructions

The MVAPICH2 installation process is designed to enable the most widely utilized features on the target build OS by default. Supported operating systems include Linux and Solaris. The default interface is OFA-IB-CH3/OFA-IWARP-CH3 on Linux and uDAPL on Solaris. The other interfaces, as indicated in Figure 1, can also be selected on Linux. This installation section provides generic instructions for building from a tarball or our latest sources.

In order to obtain best performance and scalability while having flexibility to use a large number of features, the MVAPICH team strongly recommends the use of following interfaces for different adapters: 1) OFA-IB-CH3 interface for all Mellanox InfiniBand adapters, 2) PSM-CH3 interface for all QLogic InfiniBand adapters, 3) OFA-RoCE-CH3 interface for all RoCE adapters, 4) OFA-iWARP-CH3 for all iWARP adapters and 5) Shared-Memory-CH3 for single node SMP system and laptop.

Please see the appropriate subsection for specific configuration instructions for the interface-adapter you are targeting.

4.1 Building from a tarball

The MVAPICH2 1.9 source code package includes MPICH2 3.0.3. All the required files are present as a single tarball. Download the most recent distribution tarball from:
 http://mvapich.cse.ohio-state.edu/download/mvapich2

Unpack the tarball and use the standard GNU procedure to compile:

$ tar -xzf mvapich2-1.9.tgz
$ cd mvapich2-1.9
$ ./configure
$ make
$ make install

We now support parallel make and you can use the -j<num threads> option to speed up the build process. You can use the following example to spawn 4 threads instead of the preceding make step.

$ make -j 4

In order to install a debug build, please use the following configuration option. Please note that using debug builds may impact performance.

$ ./configure --enable-g=all --enable-error-messages=all \
$ make
$ make install

4.2 Obtaining and Building the Source from SVN repository

These instructions assume you have already installed subversion.

The MVAPICH2 SVN repository is available at:  https://mvapich.cse.ohio-state.edu/svn/mpi/mvapich2

Please keep in mind the following guidelines before deciding which version to check out:

The mvapich2 directory under your present working directory contains a working copy of the MVAPICH2 source code. Now that you have obtained a copy of the source code, you need to update the files in the source tree:

$ cd mvapich2
$ ./autogen.sh

This script will generate all of the source and configuration files you need to build MVAPICH2. You will need autoconf version >= 2.67, automake version >= 1.12.3, libtool version >= 2.4

$ ./configure
$ make
$ make install

4.3 Selecting a Process Manager

MVAPICH2 provides the mpirun_rsh/mpispawn framework from MVAPICH distribution. Using mpirun_rsh should provide the fastest startup of your MPI jobs. More details can be found in Section 5.2.1. In addition, MVAPICH2 also includes the Hydra process manager from MPICH2-1.3. For more details on using Hydra, please refer to Section 5.2.2.

By default, mpiexec uses the Hydra process launcher. Please note that neither mpirun_rsh, nor Hydra require you to start daemons in advance on the nodes used for a MPI job. Both mpirun_rsh and Hydra can be used with any of the eight interfaces of this MVAPICH2 release, as indicated in Figure 1.

4.3.1 Customizing Commands Used by mpirun_rsh

Usage: ./configure [OPTION]... [VAR=VALUE]...

To assign environment variables (e.g., CC, CFLAGS...), specify them as VAR=VALUE. See below for descriptions of some of the useful variables.

RSH_CMD path to rsh command
SSH_CMD path to ssh command
ENV_CMD path to env command
DBG_CMD path to debugger command
XTERM_CMD path to xterm command
SHELL_CMD path to shell command
TOTALVIEW_CMDpath to totalview command

4.3.2 Using SLURM

If you’d like to use slurm to launch your MPI programs please use the following configure options.

$ ./configure --with-pm=no --with-pmi=slurm

4.4 Configuring a build for OFA-IB-CH3/OFA-iWARP-CH3/OFA-RoCE-CH3

OpenFabrics (OFA) IB/iWARP/RoCE with the CH3 channel is the default interface on Linux. It can be explicitly selected by configuring with:

$ ./configure --with-device=ch3:mrail --with-rdma=gen2

In order to use shared libraries configure as follows:

$ ./configure --with-device=ch3:mrail --with-rdma=gen2 --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:mrail --with-rdma=gen2 --enable-shared  --enable-g=dbg --enable-debuginfo

Configuration Options for OpenFabrics IB/iWARP/RoCE

4.5 Configuring a build for NVIDIA GPU with OFA-IB-CH3

This section details the configuration option to enable GPU-GPU communication with the OFA-IB-CH3 interface of the MVAPICH2 MPI library. For more options on configuring the OFA-IB-CH3 interface, please refer to Section 4.4.

The CUDA installation is automatically detected if installed in the standard location. To specify an alternative path to the CUDA installation, you can either use:
--with-cuda=<path/to/cuda/installation>
or
--with-cuda-include=<path/to/cuda/include>
--with-cuda-libpath=<path/to/cuda/libraries>

Note: If using the PGI compiler, you will need to add some CPPFLAGS and --enable-cuda=basic configure option to build properly. See the example below.

Example:
./configure --enable-cuda=basic CPPFLAGS="-D__x86_64
-D__align__\(n\)=__attribute__\(\(aligned\(n\)\)\)
-D__location__\(a\)=__annotate__\(a\)
-DCUDARTAPI="

4.6 Configuring a build for Shared-Memory-CH3

The default CH3 channel provides native support for shared memory communication on stand alone multi-core nodes that are not equipped with InfiniBand adapters. The steps to configure CH3 channel explicitly can be found in Section  4.4. Dynamic Process Management (5.2.5) and CheckPoint/Restart (6.15.1) are currently not supported on stand-alone nodes without InfiniBand adapters.

4.7 Configuring a build for OFA-IB-Nemesis

The Nemesis sub-channel now is supported over OFA-IB. It can be built with:

$ ./configure --with-device=ch3:nemesis:ib

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:nemesis:ib --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:nemesis:ib --enable-shared  --enable-g=dbg --enable-debuginfo

Configuration options for OFA-IB-Nemesis:

4.8 Configuring a build for uDAPL-CH3

The uDAPL interface is the default on Solaris. It can be explicitly selected on both Solaris and Linux by configuring with:

$ ./configure --with-device=ch3:mrail --with-rdma=udapl

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:mrail --with-rdma=udapl --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:mrail --with-rdma=udapl --enable-shared  --enable-g=dbg --enable-debuginfo

Configuration options for uDAPL-CH3:

4.9 Configuring a build for QLogic PSM-CH3

The QLogic PSM interface needs to be built to use MVAPICH2 on InfiniPath adapters. It can built with:

$ ./configure --with-device=ch3:psm

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:psm --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:psm --enable-shared  --enable-g=dbg --enable-debuginfo

Configuration options for QLogic PSM channel:

4.10 Configuring a build for TCP/IP-Nemesis

The use of TCP/IP with Nemesis channel requires the following configuration:

$ ./configure --with-device=ch3:nemesis

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:nemesis --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:nemesis --enable-shared  --enable-g=dbg --enable-debuginfo

Additional instructions for configuring with TCP/IP-Nemesis can be found in the MPICH2 documentation available at:  http://www.mcs.anl.gov/research/projects/mpich2/documentation/index.php?s=docs

4.11 Configuring a build for TCP/IP-CH3

The use of TCP/IP requires the explicit selection of a TCP/IP enabled channel. The recommended channel is ch3:sock and it can be selected by configuring with:

$ ./configure --with-device=ch3:sock

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:sock --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:sock --enable-shared  --enable-g=dbg --enable-debuginfo

Additional instructions for configuring with TCP/IP can be found in the MPICH2 documentation available at:

 http://www.mcs.anl.gov/research/projects/mpich2/documentation/index.php?s=docs

4.12 Configuring a build for OFA-IB-Nemesis and TCP/IP Nemesis (unified binary)

MVAPICH2 supports a unified binary for both OFA and TCP/IP communication through the Nemesis interface.

In order to configure MVAPICH2 for unified binary support, perform the following steps:

$ ./configure --with-device=ch3:nemesis:ib,tcp

You can use mpicc as usual to compile MPI applications. In order to run your application on OFA:

$ mpiexec -f hosts ./a.out -n 2

To run your application on TCP/IP:

$ MPICH_NEMESIS_NETMOD=tcp mpiexec -f hosts ./osu_latency -n 2

4.13 Configuring a build for Shared-Memory-Nemesis

The use of Nemesis shared memory channel requires the following configuration.

$ ./configure --with-device=ch3:nemesis

In order to use shared libraries, configure as follows:

$ ./configure --with-device=ch3:nemesis --enable-shared

To enable use of the TotalView debugger, the library needs to be configured in the following manner:

$ ./configure --with-device=ch3:nemesis --enable-shared  --enable-g=dbg --enable-debuginfo

Additional instructions for configuring with Shared-Memory-Nemesis can be found in the MPICH2 documentation available at:  http://www.mcs.anl.gov/research/projects/mpich2/documentation/index.php?s=docs

5 Basic Usage Instructions

5.1 Compile Applications

MVAPICH2 provides a variety of MPI compilers to support applications written in different programming languages. Please use mpicc, mpif77, mpiCC, or mpif90 to compile applications. The correct compiler should be selected depending upon the programming language of your MPI application.

These compilers are available in the MVAPICH2_HOME/bin directory. MVAPICH2 installation directory can also be specified by modifying $PREFIX, then all the above compilers will also be present in the $PREFIX/bin directory.

5.2 Run Applications

This section provides instructions on how to run applications with MVAPICH2. Please note that on new multi-core architectures, process-to-core placement has an impact on performance. Please refer to Section 6.5 to learn about running MVAPICH2 library on multi-core nodes.

5.2.1 Run using mpirun_rsh

The MVAPICH team suggests users using this mode of job start-up for all interfaces (including OFA-IB-CH3, OFA-IB-Nemesis, OFA-iWARP-CH3, OFA-RoCE-CH3, PSM-CH3, uDAPL-CH3, Shared memory-CH3, TCP/IP-CH3 and TCP/IP-Nemesis) This mpirun_rsh scheme provides fast and scalable job start-up. It scales to multi-thousand node clusters.

Prerequisites:

Examples of running programs using mpirun_rsh:

$ mpirun_rsh -np 4 n0 n0 n1 n1 ./cpi

This command launches cpi on nodes n0 and n1, two processes per node. By default ssh is used.

$ mpirun_rsh -rsh -np 4 n0 n0 n1 n1 ./cpi

This command launches cpi on nodes n0 and n1, two processes per each node using rsh instead of ssh.

MPIRUN_RSH Hostfile:

$ mpirun_rsh -np 4 -hostfile hosts ./cpi

A list of target nodes must be provided in the file hosts one per line. MPI ranks are assigned in order of the hosts listed in the hosts file or in the order they are passed to mpirun_rsh. i.e., if the nodes are listed as n0 n1 n0 n1, then n0 will have two processes, rank 0 and rank 2; whereas n1 will have rank 1 and 3. This rank distribution is known as “cyclic”. If the nodes are listed as n0 n0 n1 n1, then n0 will have ranks 0 and 1; whereas n1 will have ranks 2 and 3. This rank distribution is known as “block”.

Hostfile Format
The mpirun_rsh hostfile format allows for users to specify hostnames, one per line, optionally with a multiplier, and HCA specification.

The multiplier allows you to save typing by allowing you to specify blocked distribution of MPI ranks using one line per hostname. The HCA specification allows you to force an MPI rank to use a particular HCA.

The optional components are delimited by a ‘:’. Comments and empty lines are also allowed. Comments start with ‘#’ and continue to the next newline.

Sample hostfile

$ cat hosts  
# sample hostfile for mpirun_rsh  
host1           # rank 0 will be placed on host1  
host2:2         # rank 1 and 2 will be placed on host 2  
host3:hca1      # rank 3 will be on host3 and will used hca1  
host4:4:hca2    # ranks 4 through 7 will be on host4 and use hca2  
 
# if the number of processes specified for this job is greater than 8  
# then the additional ranks will be assigned to the hosts in a cyclic  
# fashion.  For example, rank 8 will be on host1 and ranks 9 and 10 will  
# be on host2.

Specifying Environmental Variables
Many parameters of the MPI library can be configured at run-time using environmental variables. In order to pass any environment variable to the application, simply put the variable names and values just before the executable name, like in the following example:

$ mpirun_rsh -np 4 -hostfile hosts ENV1=value ENV2=value ./cpi

Note that the environmental variables should be put immediately before the executable.

Alternatively, you may also place environmental variables in your shell environment (e.g. .bashrc). These will be automatically picked up when the application starts executing.

Note that mpirun_rsh is sensitive to the ordering of the command-line arguments.

There are many different parameters which could be used to improve the performance of applications depending upon their requirements from the MPI library. For a discussion on how to identify such parameters, see Section 8.

Job Launch using MPMD
The mpirun_rsh framework also supports job launching using MPMD mode. It permits the use of heterogeneous jobs using multiple executables and command line arguments. The following format needs to be used:

$ mpirun_rsh -config configfile -hostfile hosts

A list of different group of executables must be provided to the job launcher in the file configfile, one per line. The configfile can contain comments. Lines beginning with “#” are considered comments.

For example:

#Config file example

#Launch 4 copies of exe1 with arguments arg1 and arg2

-n 4 : exe1 arg1 arg2

#Launch 2 copies of exe2

-n 2 : exe2

A list of target nodes must be provided in the file hosts one per line and the allocation policy previously described is used.

Please note that this section only gives general information on how to run applications using mpirun_rsh. Please refer to the following sections for more information on how to run the application over various interfaces such as iWARP and RoCE.

Other Options
Other options of mpirun_rsh can be obtained using

$ mpirun_rsh --help

5.2.2 Run using Hydra (mpiexec)

Hydra is the default process manager for MPICH2. MVAPICH2 also distributes Hydra along with with mpirun_rsh. Hydra can be used either by using mpiexec or mpiexec.hydra. All interfaces of MVAPICH2 will work using Hydra. The following is an example of running a program using it:

$ mpiexec -f hosts -n 2 ./cpi

The Hydra process manager can be used to launch MPMD jobs. For example the following command:

$ mpiexec -genv FOO=1 -env BAR=1 -n 2 ./exec1 : -env BAR=2 -n 2 ./exec2

The environment variable FOO=1 passed to “-genv” is applied the environment to all executables (i.e. exec1 and exec2). The values BAR=1 applies to exec1 and BAR=2 applies to only exec2.

This process manager has many features. Please refer to the following web page for more details.

http://wiki.mcs.anl.gov/mpich2/index.php/Using_the_Hydra_Process_Manager

5.2.3 Run using SLURM

SLURM is an open-source resource manager designed by Lawrence Livermore National Laboratory. SLURM software package and its related documents can be downloaded from: http://www.llnl.gov/linux/slurm

Once SLURM is installed and the daemons are started, applications compiled with MVAPICH2 can be launched by SLURM, e.g.

$ srun -n 2 ./a.out

The use of SLURM enables many good features such as explicit CPU and memory binding. For example, if you have two processes and want to bind the first process to CPU 0 and Memory 0, and the second process to CPU 4 and Memory 1, then it can be achieved by:

$ srun --cpu_bind=v,map_cpu:0,4 --mem_bind=v,map_mem:0,1 -n2 --mpi=none ./a.out

For more information about SLURM and its features please visit SLURM website.

5.2.4 Run on PBS/Torque Clusters

You can use MVAPICH2 for clusters administered by PBS/Torque. If you are a cluster user (not an administrator), please ask your cluster administrator to install the OSC mpiexec. If you are a cluster administrator, please follow the instructions below.

You will need to download mpiexec from Ohio Supercomputer Center (OSC) at the following link.

Please note that this mpiexec is different from the mpiexec provided within MPICH2. Also note that you do not need to use either mpirun_rsh or mpiexec.hydra on a cluster that is administered with PBS/Torque. You may also choose to remove the MVAPICH2 mpiexecs from the install path to shield your users from making a mistake of trying to run MPI jobs with a wrong launcher.

Usage of OSC mpiexec is simple. An example of using this launcher is as follows:

$ /path/to/osc/mpiexec -np 4 ./a.out

5.2.5 Run with Dynamic Process Management support

MVAPICH2 (OFA-IB-CH3 interface) provides MPI-2 dynamic process management. This feature allows MPI applications to spawn new MPI processes according to MPI-2 semantics. The following commands provide an example of how to run your application.

Please refer to Section 11.75 for information about the MV2_SUPPORT_DPM environment variable.

5.2.6 Run with mpirun_rsh using OFA-iWARP Interface

The MVAPICH2 library can automatically detect iWARP cards and use them with the appropriate settings at run time. This feature deprecates the use of the environment variable MV2_USE_IWARP_MODE which was being used earlier to enable the use of iWARP devices at run time.

All the systems to be used need the following one time setup for enabling RDMA CM usage.

Programs can be executed as follows:

$ mpirun_rsh -np 2 n0 n1 prog

The iWARP interface also provides TotalView debugging and shared library support. Please refer to Section 4.4.

5.2.7 Run with mpirun_rsh using OFA-RoCE Interface

RDMA over Converged Ethernet (RoCE) is supported with the use of the run time environment variable MV2_USE_RoCE.

Programs can be executed as follows:

$ mpirun_rsh -np 2 MV2_USE_RoCE=1 prog

RoCE requires loss-less Ethernet fabric. This requires to configure Ethernet switch to treat RoCE traffic as loss-less. A separate VLAN interface needs to be created on RoCE NIC on all compute nodes and assign a private IP address

In loss-less fabric setup, MVAPICH2 can be run in RoCE mode in following two ways

5.2.8 Run using Hydra (mpiexec) with uDAPL-CH3 Interface

MVAPICH2 can be configured with the uDAPL interface, as described in the Section 4.8 . To compile MPI applications, please refer to the Section 5.1. In order to run MPI applications with uDAPL support, please specify the environmental variable MV2_DAPL_PROVIDER. As an example,

$ mpiexec -n 4 -f hosts -env MV2_DAPL_PROVIDER OpenIB-cma ./cpi

or:

$ export MV2_DAPL_PROVIDER=OpenIB-cma

$ mpiexec -n 4 -f hosts ./cpi

Please check the /etc/dat.conf file on Linux or /etc/dat/dat.conf on Solaris to find all the available uDAPL service providers. The default value for the uDAPL provider will be chosen, if no environment variable is provided at run time. If you are using OpenFabrics software stack on Linux, the default DAPL provider is OpenIB-cma for DAPL-1.2, and ofa-v2-ib0 for DAPL-2.0. If you are using Solaris, the default DAPL provider is ibd0.

The uDAPL interface also provides TotalView debugging and shared library support. Please refer to section 4.8 for more information. Please note that mpirun_rsh can also be used for running uDAPL-CH3 interface using the same MV2_DAPL_PROVIDER environment variable.

5.2.9 Run using IPoIB with mpirun_rsh or mpiexec

You would like to run an MPI job using IPoIB but your IB card is not the default interface for IP traffic. Assume that you have a cluster setup as the following:


#hostname Eth Addr IPoIB Addr
    
compute1 192.168.0.1 192.168.1.1
compute2 192.168.0.2 192.168.1.2
compute3 192.168.0.3 192.168.1.3
compute4 192.168.0.4 192.168.1.4

The Ethernet addresses are assigned to eth0 and the IPoIB addresses are assigned to the ib0 interface. The host names resolve to the 192.168.0.* addresses.

The most direct way to use the IPoIB network is to populate your hosts file with the IP addresses of the ib0 interfaces.

Example:

$ cat - > hosts
192.168.1.1
192.168.1.2
192.168.1.3
192.168.1.4

$ mpirun_rsh -hostfile hosts -n 4 ./app1

or

$ mpiexec -f hosts -n 4 ./app1

Another way to achieve this is to use the -iface option of hydra. This allows you to have your hosts file to use the host names even though they resolve to the eth0 interface.

Example:

$ cat - > hosts
compute1
compute2
compute3
compute4

$ mpiexec -f hosts -iface ib0 -n 4 ./app1

More information can be found at the following link.

5.2.10 Run using ADIO driver for Lustre

MVAPICH2 contains optimized Lustre ADIO support for the OFA-IB-CH3 interface. The Lustre directory should be mounted on all nodes on which MVAPICH2 processes will be running. Compile MVAPICH2 with ADIO support for Lustre as described in Section 4. If your Lustre mount is /mnt/datafs on nodes n0 and n1, on node n0, you can compile and run your program as follows:

$ mpicc -o perf romio/test/perf.c
$ mpirun_rsh -np 2 n0 n1 <path to perf>/perf -fname /mnt/datafs/testfile

If you have enabled support for multiple file systems, append the prefix “lustre:” to the name of the file. For example:

$ mpicc -o perf romio/test/perf.c
$ mpirun_rsh -np 2 n0 n1 ./perf -fname lustre:/mnt/datafs/testfile

5.2.11 Run using TotalView Debugger Support

MVAPICH2 provides TotalView support. The following commands provide an example of how to build and run your application with TotalView support. Note: running TotalView requires correct setup in your environment, if you encounter any problem with your setup, please check with your system administrator for help.

5.2.12 Run using a profiling library

All MPI2-functions of MVAPICH2 support the MPI profiling interface. This allows MVAPICH2 to be used by a variety of profiling libraries for MPI applications.

Two use of profiling libraries will be describe below, mpiP and Scalasca;

6 Advanced Usage Instructions

In this section, we present the usage instructions for advanced features provided by MVAPICH2.

6.1 Running on Customized Environments

In MVAPICH2 1.9, run-time variables are used to switch various optimization schemes on and off. Following is a list of optimizations schemes and the control environmental variables, for a full list please refer to the section 11:

6.2 Export Environment

Traditionally with mpirun_rsh you have to specify all environment variables that you want visible to the remote MPI processes on the command line. With the new –export option of mpirun_rsh this is no longer necessary.

6.2.1 Sample Use

Traditional Method

$ mpirun_rsh -n 2 compute1 compute2 FOO=1 BAR=baz ./osu_latency

With export option

$ export FOO=1 BAR=baz  
$ mpirun_rsh -export -n 2 compute1 compute2 ./osu_latency

6.3 Configuration File Processing

MVAPICH2 supports the use of configuration values to ease the burden of users when they would like to set and repeatedly use many different environment variables. These can be stored in a configuration file with statements of the form “VARIABLE = VALUE”. Full line comments are supported and begin with the “#” character.

The system configuration file can be placed at /etc/mvapich2.conf while user configuration files are located at “~/.mvapich2.conf” by default. The user configuration file can be specified at runtime by MV2_USER_CONFIG if the user would like to have mvapich2 read from a different location.

The processing of these files can be disabled by the use of the MV2_IGNORE_SYSTEM_CONFIG and MV2_IGNORE_USER_CONFIG.

6.3.1 Sample Use

Run with blocking mode enabled

$ cat ~/.mvapich2.conf  
# Enable blocking mode  
MV2_USE_BLOCKING = 1  
$ mpirun_rsh -n 2 compute1 compute2 ./osu_latency

Do not use user configuration file

$ mpirun_rsh -n 2 compute1 compute2 MV2_IGNORE_USER_CONFIG=1 ./osu_latency

6.4 Suspend/Resume Support

MVAPICH2 can be suspended and resumed when using a process launcher that catches and forwards the appropriate signals.

For example, when using mpirun_rsh you can type Ctrl-Z (or send the SIGTSTP signal) at the terminal and the job will suspend. You can then later send the SIGCONT signal to the job and it will continue.

6.5 Running with Efficient CPU (Core) Mapping

MVAPICH2-CH3 interfaces support architecture specific CPU mapping through the Portable Hardware Locality (hwloc) software package. By default, the HWLOC sources are compiled and built while the MVAPICH2 library is being installed. Users can choose the “–disable-hwloc” parameter while configuring the library if they do not wish to have the HWLOC library installed. However, in such cases, the MVAPICH2 library will not be able to perform any affinity related operations.

There are two major schemes as indicated below. To take advantage of any of these schemes, the jobs need to run with CPU affinity (MV2_ENABLE_AFFINITY) and shared memory
(MV2_USE_SHARED_MEM) turned on (default). If users choose to set these run-time parameters to 0, then the kernel takes care of mapping processes to cores and none of these schemes will be enabled.

To report the process mapping, users can set the environment variable MV2_SHOW_CPU_BINDING to 1 (Section  11.16).

6.5.1 Using HWLOC for CPU Mapping

Under this scheme, the HWLOC tool will be used at job-launch time to detect the processor’s micro-architecture, and then generate a suitable cpu mapping string based. Two policies are currently implemented: “bunch” and “scatter”. By default, we choose to use the “bunch” mapping. However, we also allow users to choose a binding policy through the run-time variable, MV2_CPU_BINDING_POLICY. (Section 11.14)

For example, if you want to run 4 processes per node and utilize “bunch” policy on each node, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=bunch ./a.out

The CPU binding will be set as shown in Figure 2.


PIC

Figure 2: Process placement with “bunch” CPU binding policy


If you want to run 4 processes per node and utilize “scatter” policy on each node, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=scatter ./a.out

The CPU binding will be set as shown in Figure 3.


PIC

Figure 3: Process placement with “Scatter” CPU binding policy


If two applications with four processes each need to share a given node (with eight cores) at the same time with “bunch” policy, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=bunch ./a.out

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=bunch ./b.out

The CPU binding will be set as shown in Figure 4.


PIC

Figure 4: Process placement with two applications using the “bunch” CPU binding policy


If two applications with four processes each need to share a given node (with eight cores) at the same time with “scatter” policy, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=scatter ./a.out

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_POLICY=scatter ./b.out

The CPU binding will be set as shown in Figure 5.


PIC

Figure 5: Process placement with two applications using the “scatter” CPU binding policy


The aforementioned binding is based on the core level, meaning each MPI process will be bound to a specific core. Actually, we provide different process binding level. There are three binding levels: “core”, “socket”, and “numanode” (which is designed for some multicore processor with NUMA node unit). We use the “core” as the default binding level, and we also allow users to choose a binding level through the run-time variable, MV2_CPU_BINDING_LEVEL. (Section 11.15) For example, if you want to run 4 processes per node and utilize “socket” as the binding level on each node, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_LEVEL=socket ./a.out

The CPU binding will be set as shown in Figure 6. Note: because we use “bunch” as the default binding policy, all four processes will be bound to the first socket and each of them can use all four cores in this socket. When the binding policy is “bunch” and the binding level is “socket”, processes will be bound to the same socket until the process number is larger than the core number in the socket.


PIC

Figure 6: Process placement with the “bunch” CPU binding policy and “socket” binding level


If you want to run 4 processes per node, utilize “socket” as the binding level and “scatter” as the binding policy, you can specify:

$ mpirun_rsh -np 4 -hostfile hosts MV2_CPU_BINDING_LEVEL=socket MV2_CPU_BINDING_POLICY=scatter ./a.out

The CPU binding will be set as shown in Figure 7.


PIC

Figure 7: Process placement with the “scatter” CPU binding policy and “socket” binding level


6.5.2 User defined CPU Mapping

Under the second scheme, users can also use their own mapping to bind processes to CPU’s on modern multi-core systems. The feature is especially useful on multi-core systems, where performance may be different if processes are mapped to different cores. The mapping can be specified by setting the environment variable MV2_CPU_MAPPING (Section  11.13).

For example, if you want to run 4 processes per node and utilize cores 0, 1, 4, 5 on each node, you can specify:

$ mpirun_rsh -np 64 -hostfile hosts MV2_CPU_MAPPING=0:1:4:5 ./a.out

or

$ mpiexec -n 64 -f hosts -env MV2_CPU_MAPPING 0:1:4:5 ./a.out

In this way, process 0 on each node will be mapped to core 0, process 1 will be mapped to core 1, process 2 will be mapped to core 4, and process 3 will be mapped to core 5. For each process, the mapping is separated by a single “:”.

MVAPICH2 supports binding one process to multiple cores in the same node with “,” or “-”. For example:

$ mpirun_rsh -np 64 -hostfile hosts MV2_CPU_MAPPING=0,2,3,4:1:5:6 ./a.out

or

$ mpirun_rsh -np 64 -hostfile hosts MV2_CPU_MAPPING=0,2-4:1:5:6 ./a.out

In this way, process 0 on each node will be mapped to core 0, core 2, core 3, and core 4; process 1 will be mapped to core 1, process 2 will be mapped to core 5, and process 3 will be mapped to core 6. This feature is designed to support the case that one rank process will spawn multiple threads and set thread binding in the program.

6.5.3 Performance Impact of CPU Mapping

Here we provide a table with latency performance of 0 byte and 8KB messages using different CPU mapping schemes. The results show how process binding can affect the benchmark performance. We strongly suggest the consideration of best CPU mapping on multi-core platforms when carrying out benchmarking and performance evaluation with MVAPICH2.

The following measurements were taken on the machine with the dual quad-core 2.53GHz Intel Xeon processors with 12MB L3 shared cache (among cores in one socket). MVAPICH2-1.9 was built with gcc-4.4.6 and default configure arguments:






Message Latency


Core Pair0-byte 8k-byte
Notes




1,2 0.17 us1.83 us same socket, shared L3 cache, best performance




0,1 0.17 us1.87 us same socket, shared L3 cache, but core 0 handles interrupts




1,5 0.41 us3.16 us different sockets




0,4 0.42 us3.17 us different sockets, but core 0 handles interrupts





6.6 Running with LiMIC2

MVAPICH2 CH3-based interfaces support LiMIC2 for intra-node communication for medium and large messages to get higher performance. LiMIC2 is also used to optimize intra-node one-sided communication in OFA-IB-CH3 and OFA-iWARP-CH3 interfaces. It is disabled by default because it depends on the LiMIC2 package to be previously installed. As a convenience we have distributed the latest LiMIC2 package (as of this release) with our sources.

To install this package, please take the following steps.

Before using LiMIC2 you’ll need to load the kernel module. If you followed the instructions above you can do this using the following command (LSB init script).

Please note that supplying ‘--sysconfdir=/etc’ in the configure line above told the package to install the init script and an udev rule in the standard location for system packages. Supplying ‘--prefix=/usr’ will also install the headers and libraries in the system path. These are optional but recommended.

Now you can use LiMIC2 with MVAPICH2 by simply supplying the ‘--with-limic2’ option when configuring MVAPICH2. You can run your applications as normal and LiMIC2 will be used by default. To disable it at run time, use the env variable:

$ mpirun_rsh -np 64 -hostfile hosts MV2_SMP_USE_LIMIC2=0 ./a.out

LiMIC2 optimizations for one-sided communication are currently available for Put and Get operations with both active and passive synchronization semantics. Using LIMIC2, MVAPICH2 provides truly passive synchronization within a node. These optimizations are enabled by default, when MVAPICH2 is configured with “–with-limic2”. They can be disabled at run time, as below:

$ mpirun_rsh -np 64 -hostfile hosts MV2_USE_LIMIC_ONE_SIDED=0 ./a.out

6.7 Running with Shared Memory Based Windows for One-sided Communication

MVAPICH2 optimizes one-sided communication when the windows are created in shared memory. Programmers can use the MPI_Info argument in MPI_Alloc_mem to provide a hint (key: alloc_shm, value: true) when the buffer will be used to create a window and should be allocated in shared memory. MVAPICH2 internally detects the windows created on such buffers and optimizes the communication on them. The snippet below shows how a user can provide the hint in a MPI_Alloc_mem call. One-sided communication benchmarks in the OMB suite serve as complete examples. Through this design MVAPICH2 provides truly passive one-sided semantics within a node.

MPI_Info_create(&win_info); MPI_Info_set(win_info, ”alloc_shm”, ”true”); MPI_Alloc_mem (size, win_info, &buffer);

Users who want to take advantage of this design without changing the application code (to provide the hint) can set the run time parameter MV2_USE_SHM_ONE_SIDED ( 11.94) to 1. When this is set, all the buffers allocated using MPI_Alloc_mem are allocated in shared memory by default. This helps optimize the one-sided communication that happens on this memory. However, users have to note that the buffers still have to be allocated using MPI_Alloc_mem and this option is less scalable as the amount of shared memory allocated on a node can be limited. Hence, use of the optimization hint described above is recommended.

6.8 Running with Shared Memory Collectives

In MVAPICH2, support for shared memory based collectives has been enabled for MPI applications running over OFA-IB-CH3, OFA-iWARP-CH3, uDAPL-CH3 and PSM-CH3 interfaces. Currently, this support is available for the following collective operations:

Optionally, these feature can be turned off at run time by using the following parameters:

Please refer to Section 11 for further details.

6.9 Running Collectives with Hardware based Multicast support

In MVAPICH2, support for multicast based collectives has been enabled for MPI applications running over OFA-IB-CH3 interface. Currently, this support is available for the following collective operations:

This feature is disabled by default. This can be turned on at runtime by using parameter MV2_USE_MCAST ( 11.124). This feature is effective when the MPI job is running on more than the threshold
MV2_MCAST_NUM_NODES_THRESHOLD ( 11.125) number of nodes.

This feature requires the cluster to be installed with libibumad and libibmad libraries. If there are not installed, this feature can be disabled with –disable-mcast configure flag.

This feature requires to have read/write permission for users on /dev/infiniband/umad0

$ ls -l /dev/infiniband/umad0  
crw-rw-rw- 1 root root 231, 0 Jul  8 19:47 /dev/infiniband/umad0

6.10 Running MPI_Gather collective with intra-node Zero-Copy designs (using LiMIC2)

In MVAPICH2, we offer new intra-node Zero-Copy designs (using LiMIC2) for the MPI_Gather collective operation based on the LiMIC2 feature. This feature can be used, when the library has been configured to use LiMIC2( 6.6). This feature is disabled by default and can be turned on at runtime by using the parameter MV2_USE_LIMIC_GATHER ( 11.123).

6.11 Running with scalable UD transport

MVAPICH2 has scalable design with Infiniband connection less transport Unreliable Datagram (UD). Applications can use UD only transport by simply configuring MVAPICH2 with the –enable-hybrid and setting the environment variable MV2_USE_ONLY_UD( 11.121). In this mode, library not uses any reliable RC connections. This feature eliminates all the overheads associated with RC connections and reduces the memory footprint at large scale.

6.12 Running with Integrated Hybrid UD-RC/XRC design

MVAPICH2 has integrated hybrid transport support for OFA-IB-CH3. This provides the capability to use Unreliable Datagram (UD), Reliable Connection (RC) and eXtended Reliable Connection (XRC) transports of InfiniBand. This hybrid transport design is targeted at emerging clusters with multi-thousand core clusters to deliver best possible performance and scalability with constant memory footprint.

Applications can use Hybrid transport by simply configuring MVAPICH2 with the –enable-hybrid option. In this configuration, MVAPICH2 seamlessly uses UD and RC/XRC connections by default. The use of UD transport can be disabled at run time by setting the environment variable MV2_USE_UD_HYBRID( 11.120) to Zero.

MV2_HYBRID_ENABLE_THRESHOLD ( 11.115) defines the threshold for enabling the hybrid transport. Hybrid mode will be used when the size of the job is greater than or equal to the threshold. Otherwise, it uses default RC/XRC connections.

For a full list of Hybrid environment variables, please refer Section 11.

6.13 Running with Multiple-Rail Configurations

MVAPICH2 has integrated multi-rail support for OFA-IB-CH3 and OFA-iWARP-CH3 interfaces. Run-time variables are used to specify the control parameters of the multi-rail support; number of adapters with MV2_NUM_HCAS (section 11.35), number of ports per adapter with MV2_NUM_PORTS (section 11.36), and number of queue pairs per port with MV2_NUM_QP_PER_PORT (section 11.38). Those variables are default to 1 if you do not specify them.

Large messages are striped across all HCAs. The threshold for striping is set according to the following formula:
(MV2_VBUF_TOTAL_SIZE × MV2_NUM_PORTS × MV2_NUM_QP_PER_PORT × MV2_NUM_HCAS). In addition, there is another parameter MV2_STRIPING_THRESHOLD (section 11.74) which users can utilize to set the striping threshold directly.

MVAPICH2 also gives the flexibility to balance short message traffic over multiple HCAs in a multi-rail configuration. The run-time variable MV2_SM_SCHEDULING can be used to choose between the various load balancing options available. It can be set to USE_FIRST (Default) or ROUND_ROBIN. In the USE_FIRST scheme, the HCA in slot 0 is always used to transmit the short messages. If ROUND_ROBIN is chosen, messages are sent across all HCAs alternately.

In the following example, we can use multi-rail support with two adapters, using one port per adapter and one queue pair per port:

$ mpirun_rsh -np 2 n0 n1 MV2_NUM_HCAS=2 MV2_NUM_PORTS=1 MV2_NUM_QP_PER_PORT=1 prog

Using the Hydra process manager, the same can be accomplished by:

$ mpiexec -n 2 -hosts n0,n1 -env MV2_NUM_HCAS 2 -env MV2_NUM_PORTS 1 -env MV2_NUM_QP_PER_PORT 1 prog

Note that the default values of MV2_NUM_PORTS and MV2_NUM_QP_PER_PORT are 1, so they can be omitted.

$ mpirun_rsh -np 2 n0 n1 MV2_NUM_HCAS=2 prog

Using the Hydra process launcher, the following command can be used:

$ mpiexec -n 2 -hosts n0,n1 -env MV2_NUM_HCAS 2 prog

The user can also select the particular network card(s) that should be used by using the MV2_IBA_HCA environment variable specified in section 11.25. The following is an example of how to run MVAPICH2 in this mode. (In the example “mlx4_0” is the name of the InfiniBand card as displayed by the output of the “ibstat” command).

$ mpirun_rsh -np 2 n0 n1 MV2_IBA_HCA=mlx4_0 prog

If there are multiple HCAs in the system, the user can also selectively use some or all of these HCAs for network operations by using the MV2_IBA_HCA environment variable. Different HCAs are delimited by colons “:”. An example is shown below. In the example “mlx4_0” and “mlx4_1” are the names of the InfiniBand card as displayed by the output of the “ibstat” command. There can be other HCAs in the system as well.

$ mpirun_rsh -np 2 n0 n1 MV2_IBA_HCA=mlx4_0:mlx4_1 prog

6.14 Enhanced design for Multiple-Rail Configurations

MVAPICH2 now features an enhanced design for multiple rail configurations for OFA-IB-CH3 and OFA-iWARP-CH3 interfaces. It can broadly be explained by the figure given below. In addition to the earlier design where the rails were shared amongst processes at run time (as depicted under the Rail Sharing banner in the figure below), MVAPICH2 now features a new RAIL BINDING policy which will dedicate a particular rail to a particular process.


PIC

Figure 8: Multi-rail Scheduling policies


The scheduling policies are broadly classified into 2 basic types. Rail Binding and Rail Sharing.

6.15 Running with Fault-Tolerance Support

6.15.1 System-Level Checkpoint/Restart

MVAPICH2 provides system-level checkpoint/restart functionality based on a coordinated Checkpoint/Restart protocol involving all the application processes.

The following section ( 6.15.1) provides instructions for basic checkpoint/restart operations with MVAPICH2. These require the Berkeley Lab Checkpoint/Restart (BLCR) library in order perform checkpoints of local processes. Then, advanced features are presented. Section 6.15.1 details the usage of the fast-checkpointing scheme based on aggregation. Section 6.15.1 present the support of the new standardized Fault Tolerance Backplane (FTB).

Basic Checkpoint/Restart Scheme: BLCR is a library that allows to take checkpoint of individual processes. Its usage is mandatory to take advantage of the checkpoint/restart functionality in MVAPICH2. Here are the steps that allows the usage of BLCR with MVAPICH2.

Users are strongly encouraged to read the Administrators guide of BLCR, and test the BLCR on the target platform, before using the checkpointing feature of MVAPICH2.

Checkpointing operation

Now, your system is set up to use the Checkpoint/Restart features of MVAPICH2. Several parameters are provided by MVAPICH2 for flexibility in configuration and using the Checkpoint / Restart features. If mpiexec is used as the job start up mechanism, these parameters need to be set in the user’s environment through the BASH shell’s export command, or the equivalent command for other shells. If mpirun_rsh is used as the job start up mechanism, these parameters need to be passed to mpirun_rsh through the command line.

In order to provide maximum flexibility to end users who wish to use the checkpoint/restart features of MVAPICH2, we’ve provided three different methods which can be used to take the checkpoints during the execution of the MPI application. These methods are described as follows:

Restart operation

To restart a job from a manual checkpoint, users need to issue another command of BLCR, “cr_restart” with the checkpoint file name of the MPI job console as the parameter. Usually, this file is named
context.<pid>. The checkpoint file name of the MPI job console can be specified when issuing the checkpoint (see the “cr_checkpoint --help” for more information). Please note that the names of checkpoint files of the MPI processes will be assigned according to the environment variable MV2_CKPT_FILE,
($MV2_CKPT_FILE.<number of checkpoint>.<process rank>).

To restart a job from an automatic checkpoint, use cr_restart $MV2_CKPT_FILE.<number of checkpoint>.auto.

If the user wishes to restart the MPI job on a different set of nodes, the host file that was specified along with the “-hostfile” option during job launch phase should be modified accordingly before trying to restart a job with “cr_restart”. This modified “hostfile” must be at the same location and with the same file name as the original hostfile. The mpirun_rsh framework parses the host file again when trying to restart from a checkpoint, and launches the job on the corresponding nodes. This is possible as long as the nodes in which the user is trying to restart has the exact same environment as the one in which the checkpoint was taken (including shared NFS mounts, kernel versions, and user libraries).

For this to function correctly, the user should disable pre-linking on both the source and the destination node. See the  FAQ Section of the BLCR userguide for more information.

Please refer to the Section 9.7 for troubleshooting with Checkpoint/Restart.

Write-Aggregation based Fast Checkpointing Scheme: MVAPICH2 provides an enhanced technique that allows fast checkpoint and restart. This scheme, named Aggregation, relies on the Filesystem in Userspace (FUSE) library.

Although Aggregation is optional, its usage is recommended to achieve best performances during checkpoint and restart operations. That is why, if the FUSE library is detected during configuration step, it will be automatically enabled (see section 4.4). Once enabled at configuration step, aggregation scheme can be disabled at run time by setting the environment variable MV2_CKPT_USE_AGGREGATION=0 (see section 11.8 for details).

The following steps need to be done to use FUSE library for aggregation scheme.

If write aggregation has been enabled at configuration time, MVAPICH2 will check the FUSE configuration of each node during startup (FUSE module loaded and fusermount command in the PATH). If one node is not properly configured, then MVAPICH2 will abort. In this case, you need to fix the FUSE configuration of the nodes, or disable aggregation using MV2_CKPT_USE_AGGREGATION=0 to run MVAPICH2.

Fault Tolerance Backplane (FTB) support: MVAPICH2 supports the new standardized Fault Tolerance Backplane (FTB). FTB can be used for Checkpoint-Restart and Job Pause-Migration-Restart Frameworks. Activating FTB support is optional to perform checkpoint/restart operations, but it will allow MVAPICH2 to automatically publish and listen to standardized events through the Fault Tolerance Backplane.

FTB has been developed and standardized by the CIFTS project. It enables faults to be handled in a coordinated and holistic manner in the entire system, providing for an infrastructure which can be used by different software systems to exchange fault-related information.

If using the FTB framework for checkpoint/restart, the following steps need to be done in addition to the above nodes.

Checkpoint/Restart support for the OFA-IB-Nemesis Interface: MVAPICH2 also provides Checkpoint-Restart capabilities for the OFA-IB-Nemesis interface. Currently, this feature can only be used in conjunction with the Hydra process manager. More details on using Hydra to checkpoint an MPI application can be found here.

6.15.2 Multi-Level Checkpointing with Scalable Checkpoint-Restart (SCR)

The  Scalable Checkpoint-Restart (SCR) library developed at the Lawrence Livermore National Laboratory (LLNL) enables MPI applications to utilize distributed storage on Linux clusters to attain high file I/O bandwidth for checkpointing and restarting large-scale jobs. With SCR, jobs run more efficiently, lose less work upon a failure, and reduce load on critical shared resources such as the parallel file system and the network infrastructure.

In the current SCR implementation, application checkpoint files are cached in storage local to the compute nodes, and a redundancy scheme is applied such that the cached files can be recovered even after a failure disables part of the system. SCR supports the use of spare nodes such that it is possible to restart a job directly from its cached checkpoint, provided the redundancy scheme holds and provided there are sufficient spares.

The SCR library implements three redundancy schemes which trade of performance, storage space, and reliability:

SCR is integrated into MVAPICH2 to provide Multi-Level checkpointing capabilities to MPI applications in two modes: application-aware mode and transparent mode. The following sub-sections illustrate the steps needed to checkpoint an application using these two schemes.

Application-Aware Multi-Level Checkpointing: In this case, it is assumed that the application knows what data to checkpoint and how to read data from a checkpoint during a restart. It is also assumed that each process writes its checkpoint data in a unique file. For an application to actually write a checkpoint using SCR, the following steps need to be followed.

The following code snippet describes the checkpointing function of a sample MPI application that writes checkpoints using SCR:

/* Determine whether we need to checkpoint */  
int flag;  
SCR_Need_checkpoint(&flag);  
if (flag) {  
    /* Tell SCR that a new checkpoint is starting */  
    SCR_Start_checkpoint();  
 
    /* Define the checkpoint file name for this process */  
    int rank;  
    char name[256];  
    MPI_Comm_rank(MPI_COMM_WORLD, &rank );  
    sprintf (name, "rank_%d.ckpt", rank);  
 
    /* Register our file, and get the full path to open it */  
    char file[SCR MAX FILENAME];  
    SCR_Route_file name , file);  
 
    /* Open, write, and close the file */  
    int valid = 0;  
    FILE *fs = open( file, "w") ;  
    if ( fs != NULL ) {  
        valid = 1;  
        size_t n = fwrite(checkpoint_data, 1, sizeof(checkpoint_data), fs);  
        if (n != sizeof(checkpoint data)) {  
            valid = 0;  
        }  
        if ( fclose(fs) != 0 ) {  
            valid = 0;  
        }  
    }  
 
    /* Tell SCR that this process is done writing its checkpoint files */  
    SCR_Complete_checkpoint( valid ) ;  
}

The following code snippet describes the SCR-assisted restart mechanism for a sample MPI application:

 
/* Define the checkpoint filename for this process */  
int rank;  
char name[256];  
MPI_Comm_rank(MPI_COMM_WORLD, &rank);  
sprintf( name, "rank_%d.ckpt", rank);  
/* Get the full path to open our file */  
char file[SCR_MAX_FILENAME];  
if (SCR_Route_file( name, file) == SCR SUCCESS) {  
    /* Open, read, and close the file */  
    FILE *fs = open( file, "r");  
    size_t n = fread(checkpoint_data, 1, sizeof(checkpoint_data), fs);  
    fclose(fs);  
} else {  
    /* There is no existing file to restart from */  
}  

For more instructions on integrating the SCR APIs into an MPI application, and for information about the various runtime parameters that can be set by the user, please refer to the  SCR Userguide.

Transparent Multi-Level Checkpointing :

The benefits of SCR can also be leveraged by applications that do not employ their own Checkpoint-Restart mechanism. The SCR-MVAPICH2 integration makes this possible by using SCR to manage the checkpoints generated by BLCR the basic system-level Checkpointing scheme described in Section 6.15.1. Once MVAPICH2 has been configured with SCR, the steps required to checkpoint and restart a job transparently using BLCR and SCR are the same as what is described in Section 6.15.1. MVAPICH2 uses SCR’s APIs internally to transparently manage checkpoint storage efficiently.

For information about the various SCR-specific runtime parameters that can be set by the user, and for detailed information about the redundancy schemes employed by SCR, please refer to the  SCR Userguide.

6.15.3 Job Pause-Migration-Restart Support

MVAPICH2 provides a node-level Job Pause-Migration-Restart mechanism for the OFA-IB-CH3 interface, which can be used to migrate MPI processes on a given failing node to a healthy node, selected from a pool of spare hosts provided by the user during job-launch.

This Job Migration framework relies on BLCR and FTB libraries. See subsections in  6.15.1 to set up these libraries. The source and target nodes are required to have these libraries installed in the same path.

The Job Migration framework makes use of BLCR to take a checkpoint of all MPI processes running on the failing node in question. Users are strongly recommended to either disable “prelinking” feature, or execute the following command:

$ prelink --undo --all

on all nodes before starting a job that will later be migrated. Please refer to this BLCR web page for complete information:  BLCR FAQ.

During job-launch, the list of spare hosts can be provided as input to mpirun_rsh using the -sparehosts option which takes a hostfile as argument.

The following example illustrates a sample MPI job being launched to run with job migration support:

$ mpirun_rsh -np 4 -hostfile ./hosts -sparehosts ./spare_hosts ./prog

where spare_hosts is a file which contains a list of healthy spare hosts that is needed by the job migration framework.

An actual migration of MPI processes from a failing source node to a target spare node can be triggered using one of two methods - using signals or using the mv2_trigger utility.

For the signal-triggered method, users can manually initiate the migration protocol by issuing a SIGUSR2 signal to the mpispawn processes running on the source node. This can be done using the following command:

$ pkill -SIGUSR2 mpispawn

The migration protocol can also be triggered using the simple utility provided with MVAPICH2:

$ $PREFIX/bin/mv2_trigger src

where src is the hostname of the health-deteriorating node from which all MPI processes need to be migrated.

Please note, that the ftb_agent daemon will have to be launched on all the compute nodes and the spare nodes before launching the MPI application, in order to successfully migrate MPI processes. See section  6.15.1 for information about setting up the FTB infrastructure.

6.15.4 Run-Through Stabilization

MVAPICH2 provides support for run-through stabilization wherein communication failures are not treated as fatal errors. On enabling this capability, MVAPICH2 returns the appropriate error code to a user-set error handler in the event of a communication failure, instead of terminating the entire job. When a process detects a failure when communicating with another process, it will consider the other process as having failed and will no longer attempt to communicate with that process. The user can, however, continue making communication calls to other processes. Any outstanding send or receive operations to a failed process, or wild-card receives (i.e., with MPI_ANY_SOURCE) posted to communicators with a failed process, will be immediately completed with an appropriate error code.

Currently, this support is available only for the OFA-IB-Nemesis and TCP/IP-Nemesis interfaces, when the Hydra process manager is used. It can be enabled at run-time by:

6.15.5 Network Fault Tolerance with Automatic Path Migration

MVAPICH2 supports network fault recovery by using InfiniBand Automatic Path Migration (APM) mechanism for OFA-IB-CH3 interface. This support is available for MPI applications using OpenFabrics stack and InfiniBand adapters.

To enable this functionality, a run-time variable, MV2_USE_APM (Section 11.76) can be enabled, as shown in the following example:
$ mpirun_rsh -np 2 n0 n1 MV2_USE_APM=1 ./cpi
or
$ mpiexec -n 2 -hosts n0,n1 -env MV2_USE_APM 1 ./cpi

MVAPICH2 also supports testing Automatic Path Migration in the subnet in the absence of network faults. This can be controlled by using a run-time variable MV2_USE_APM_TEST (Section 11.77). This should be combined with MV2_USE_APM as follows:
$ mpirun_rsh -np 2 n0 n1 MV2_USE_APM=1 MV2_USE_APM_TEST=1 ./cpi
or
$ mpiexec -n 2 -hosts n0,n1 -env MV2_USE_APM 1 -env MV2_USE_APM_TEST 1 ./cpi

6.16 Running with RDMA CM support

In MVAPICH2, for using RDMA CM the run time variable MV2_USE_RDMA_CM needs to be used as described in 11. This applies to OFA-IB-CH3 and OFA-iWARP-CH3 interfaces.

In addition to these flags, all the systems to be used need the following one time setup for enabling RDMA CM usage.

Programs can be executed as follows:
$ mpirun_rsh -np 2 n0 n1 MV2_USE_RDMA_CM=1 prog
or
$ mpiexec -n 2 -hosts n0,n1 -env MV2_USE_RDMA_CM 1 prog

6.17 Running MVAPICH2 in Multi-threaded Environments

MVAPICH2 binds processes to processor cores for optimal performance. Please refer to Section 6.5 for more details. However, in multi-threaded environments, it might be desirable to have each thread compute using a separate processor core. This is especially true for OpenMP+MPI programs.

In MVAPICH2, processor core mapping is turned off in the following way to enable the application in MPI_THREAD_MULTIPLE threading level if user requested it in MPI_Init_thread. Otherwise, applications will run in MPI_THREAD_SINGLE threading level.

$ mpirun_rsh -np 2 n0 n1 MV2_ENABLE_AFFINITY=0 ./openmp+mpi_app

For QLogic PSM Interface,

$ mpirun_rsh -np 2 n0 n1 MV2_ENABLE_AFFINITY=0 IPATH_NO_CPUAFFINITY=1 ./openmp+mpi_app

6.18 Compiler Specific Flags to enable OpenMP thread binding

Further, to get better performance for applications that use MPI + OpenMP, we recommend binding the OpenMP threads to the processor cores. This can potentially avoid cache effects due to unwanted thread migration. For example, if we consider a Quad-Core processor, with 1 MPI process and 4 OpenMP threads, we recommend binding OpenMP thread 0 to core 0, OpenMP thread 1 to core 1 and so on. This can be achieved by setting the kernel affinity via the following compiler-specific environment variables:

Intel Compilers: On Intel processors, we recommend using the the KMP_AFFINITY run-time flag, in the following manner:

mpirun_rsh -hostfile hostfile -np 1 MV2_ENABLE_AFFINITY=0 OMP_NUM_THREADS=4
KMP_AFFINITY=warnings,compact ./a.out

For more information, please refer to: (http://software.intel.com/sites/products/documentation/studio/
composer/en-us/2011Update/compiler_c/optaps/common/optaps_openmp_thread_affinity.htm)

GNU compilers: On both Intel and AMD processors, we recommend using the GOMP_CPU_AFFINITY run-time flag, in the following manner:

mpirun_rsh -hostfile hostfile -np 1 MV2_ENABLE_AFFINITY=0 OMP_NUM_THREADS=4
GOMP_CPU_AFFINITY=”0,1,2,3” ./a.out

For more information, please refer to:
(http://gcc.gnu.org/onlinedocs/libgomp/GOMP_005fCPU_005fAFFINITY.html)

6.19 Running with Hot-Spot and Congestion Avoidance

MVAPICH2 supports hot-spot and congestion avoidance using InfiniBand multi-pathing mechanism. This support is available for MPI applications using OFA-IB-CH3 interface.

To enable this functionality, a run-time variable, MV2_USE_HSAM (Section 11.82) can be enabled, as shown in the following example:
$ mpirun_rsh -np 2 n0 n1 MV2_USE_HSAM=1 ./cpi
or
$ mpiexec -n 2 -hosts n0,n1 -env MV2_USE_HSAM 1 ./cpi

This functionality automatically defines the number of paths for hot-spot avoidance. Alternatively, the maximum number of paths to be used between a pair of processes can be defined by using a run-time variable MV2_NUM_QP_PER_PORT (Section 11.38).

We expect this functionality to show benefits in the presence of at least partially non-overlapping paths in the network. OpenSM, the subnet manager distributed with OpenFabrics supports LMC mechanism, which can be used to create multiple paths:

$ opensm -l4

will start the subnet manager with LMC value to four, creating sixteen paths between every pair of nodes.

6.20 Running on Clusters with NVIDIA GPU Accelerators

MVAPICH2 CH3-based interface supports MPI communication using NVIDIA GPU device memory with CUDA versions 4.0 or later. This feature removes the need for the application developer to explicitly move the data from device memory to host memory before using MPI for communication. The new support allows direct MPI communication from device memory to device memory, device memory to host memory and host memory to device memory. It also supports point-to-point and collective communication using contiguous and non-contiguous MPI datatypes. It takes advantage of CUDA IPC for intra-node GPU-GPU communication (with CUDA 4.1).

For example, without CUDA support in the MPI library, a typical user might be using the following sequence of commands to move data from a device memory to another device memory.


cudaMemcpy(host_buf, device_buf, size, cudaMemcpyDeviceToDevice);
MPI_Isend(host_buf, size, MPI_CHAR, 1, 100, MPI_COMM_WORLD, req);


With the support provided in MVAPICH2 and support of CUDA 4.0 (and later), the user can achieve the same data movement operation by explicitly specifying MPI calls on device memory.

MPI_Isend(device_buf, size, MPI_CHAR, 1, 100, MPI_COMM_WORLD, req);

This support can be enabled by configuring MVAPICH2 with --enable-cuda and setting the environment variable MV2_USE_CUDA ( 11.126) to 1 during runtime.

To minimize communication overhead, MVAPICH2 divides copies between device and host into chunks. This can be better tuned for internode transfers with a runtime environment variable MV2_CUDA_BLOCK_SIZE ( 11.127). The default chunk size is 64K (65536). However, higher values of this parameter, such as 256K (262144) and 512K (524288), might deliver better performance if the MPI application uses large messages. The optimal value for this parameter depends on several other factors such as InfiniBand network/adapter speed, GPU adapter characteristics, platform characteristics (processor and memory speed) and amount of memory to be dedicated to the MPI library with GPU support. For different platforms and MPI applications, the users are encouraged to try out different values for this parameter to get best performance. For intranode transfers, the chunk size is controlled by the runtime environment variable MV2_SMP_SEND_BUF_SIZE and its default value is 128K (131072) when CUDA support is enabled. Other recommended values for this variable are 64K (65536) and 256K (262144). Users have to note that the parameter MV2_SMP_SEND_BUF_SIZE will also impact host-to-host intranode communication in which case smaller chunk sizes (8K or 16K) deliver better performance.

GPU Affinity: When multiple GPUs are present on a node, users might want to set the MPI process affinity to a particular GPU using cuda calls like cudaSetDevice(). This can be done after MPI_Init based on MPI rank of the process. But MVAPICH2 performs some cuda operations like buffer registration and others during MPI_Init which result in default context creation. Hence, setting GPU Affinity after MPI_Init could create issues due to the context switch. To avoid this, MVAPICH2 provides an environment variable called MV2_COMM_WORLD_LOCAL_RANK to get the local rank of a process on its node before MPI_Init is called. This local rank information can be used to set GPU affinity before MPI_Init is called as given in the following code example

....
int local_rank = atoi(getenv(”MV2_COMM_WORLD_LOCAL_RANK”));
cudaSetDevice(local_rank % num_devices);
...
MPI_Init()
...


This local rank information can also be used in wrapper scripts to set cpu and memory affinity on nodes with NUMA.

6.21 MPIRUN_RSH compatibility with MPIEXEC

There is now a front end to mpirun_rsh that aims for compatibility with mpiexec usage as recommended by the MPI standard. This front end is can be used by calling mpiexec.mpirun_rsh.

The main differences between this and mpirun_rsh is that this front end doesn’t require you to specify hosts or environment variables on the command line. Also, the host file options is -f as opposed to -hostfile.

Run 4 instances of CPI on localhost:

$ mpiexec.mpirun_rsh -n 4 ./cpi

Run 4 instances of CPI on the nodes specified by hostfile:

$ mpiexec.mpirun_rsh -n 4 -f hostfile ./cpi

6.21.1 Interaction with SLURM

If the hostfile isn’t given and mpiexec.mpirun_rsh finds the appropriate slurm environment variables set, mpiexec will use these to determine which nodes the applications should be launched on.

Please note that processes are assigned in a block fashion by slurm so all the processes will run on the same node if possible.

Run 4 instances of CPI on up to 4 different nodes (but maybe only 1) allocated by “salloc”:

$ salloc -N 4
$ mpiexec.mpirun_rsh -n 4 ./cpi

Same thing as above but only allocating the minimum number of nodes needed to run 4 tasks:

$ salloc -n 4
$ mpiexec.mpirun_rsh ./cpi

Run 4 instances of CPI on 4 different nodes allocated by “salloc”:

$ salloc -N 4 --ntasks-per-node=1
$ mpiexec.mpirun_rsh -n 4 ./cpi

6.21.2 Interaction with PBS

If the hostfile isn’t given and mpiexec.mpirun_rsh finds the PBS_NODEFILE environment variable set, mpiexec will use this to locate the host file.

7 OSU Benchmarks

If you have arrived at this point, you have successfully installed MVAPICH2. Congratulations!! The OSU benchmarks should already be built and installed along with MVAPICH2. Look for them under $prefix/libexec/mvapich2. Sample performance numbers for these benchmarks on representative platforms with InfiniBand, iWARP and RoCE adapters are also included on our projects’ web page. You are welcome to compare your performance numbers with our numbers. If you see any big discrepancy, please let us know by sending an email to mvapich-discuss@cse.ohio-state.edu.

The benchmarks provided are:

MPI-1, MPI-2 and MPI-3


osu_bibw Bidirectional Bandwidth Test
osu_bw Bandwidth Test
osu_latency Latency Test
osu_mbw_mr Multiple Bandwidth / Message Rate Test
osu_multi_lat Multi-pair Latency Test
osu_allgather MPI_Allgather Latency Test
osu_allgatherv MPI_Allgatherv Latency Test
osu_allreduce MPI_Allreduce Latency Test
osu_alltoall MPI_Alltoall Latency Test
osu_alltoallv MPI_Alltoallv Latency Test
osu_barrier MPI_Barrier Latency Test
osu_bcast MPI_Bcast Latency Test
osu_gather MPI_Gather Latency Test
osu_gatherv MPI_Gatherv Latency Test
osu_reduce MPI_Reduce Latency Test
osu_reduce_scatter MPI_Reduce_scatter Latency Test
osu_scatter MPI_Scatter Latency Test
osu_scatterv MPI_Scatterv Latency Test
MPI-2 and MPI-3 only


osu_acc_latency Accumulate Latency Test with Active Synchronization
osu_get_bw One-Sided Get Bandwidth Test with Active Synchronization
osu_get_latency One-Sided Get Latency Test with Active Synchronization
osu_latency_mt Multi-threaded Latency Test
osu_passive_acc_latencyLatency Test for Accumulate with Passive Synchronization
osu_passive_get_bw Bandwidth Test for Get with Passive Synchronization
osu_passive_get_latency Latency Test for Get with Passive Synchronization
osu_passive_put_bw Bandwidth Test for Put with Passive Synchronization
osu_passive_put_latencyLatency Test for Put with Passive Synchronization
osu_put_bibw One-Sided Put Bidirectional Test with Active Synchronization
osu_put_bw One-Sided Put Bandwidth Test with Active Synchronization
osu_put_latency One-Sided Put Latency Test with Active Synchronization

More information about the benchmarks can be found at http://mvapich.cse.ohio-state.edu/benchmarks/. You can also check this link for updates to the benchmarks.

7.1 Download and Build Stand-alone OSU Benchmarks Package

The OSU Benchmarks can also be downloaded as a separate package from here. You can build the benchmarks using the following steps if mpicc is in your PATH. For example:

$ ./configure --prefix=<path-to-install> && make && make install

If mpicc is not in your path or you would like to use another particular version you can explicitly tell configure by setting CC. For example:

$ ./configure CC=/path/to/special/mpicc --prefix=<path-to-install> && make && make install

Configure will detect whether your library supports MPI-2, MPI-3 and compile the corresponding benchmarks. The benchmarks will be installed under $prefix/libexec/osu-micro-benchmarks.

CUDA Extensions to OMB can be enabled by configuring the benchmark suite with –enable-cuda option as shown below. Similarly, OpenACC Extensions can be enabled by specifying the –enable-openacc option. The MPI library used should be able to support MPI communication from buffers in GPU Device memory.

$ ./configure CC=/path/to/mpicc
    --enable-cuda
    --with-cuda-include=/path/to/cuda/include
    --with-cuda-lib=/path/to/cuda/lib
$ make
$ make install

7.2 Running

The OSU Benchmarks are run in the same manner as other MPI Applications. The following examples will use mpirun_rsh as the process manager. Please see section 5.2 for more information on running with other process managers.

7.2.1 Running OSU Latency and Bandwidth

Inter-node latency and bandwidth: The following example will measure the latency and bandwidth of communication between node1 and node2.

$ mpirun_rsh -np 2 node1 node2 ./osu_latency

$ mpirun_rsh -np 2 node1 node2 ./osu_bw

Intra-node latency and bandwidth: The following example will measure the latency and bandwidth of communication inside node1 on different cores. This assumes that you have at least two cores (or processors) in your node.

$ mpirun_rsh -np 2 node1 node1 ./osu_latency

$ mpirun_rsh -np 2 node1 node1 ./osu_bw

7.2.2 Running OSU Message Rate Benchmark

The OSU message rate benchmark reports the rate at which messages can be sent between two nodes. It is advised that it should be run in a configuration that utilizes multiple pairs of communicating processes on two nodes. The following example measures the message rate on a system with two nodes, each with four processor cores.

$ mpirun_rsh -np 8 -hostfile mf ./osu_mbw_mr

Where hostfile “mf” has the following contains:

node1
node1
node1
node1
node2
node2
node2
node2

7.2.3 Running OSU Collective Benchmarks

By default, the OSU collective benchmarks report the average communication latencies for a given collective operation, across various message lengths. Additionally, the benchmarks offer the following options:

  1. ”-f” can be used to report additional statistics, such as min and max latencies and the number of iterations.
  2. ”-m” option can be used to set the maximum message length to be used in a benchmark. In the default version, the benchmarks report the latencies for up to 1MB message lengths.
  3. ”-i” can be used to set the number of iterations to run for each message length.
  4. ”-h” can be used to list all the options and their descriptions.
  5. ”-v” reports the benchmark version.

If a user wishes to measure the communication latency of a specific collective operation, say, MPI_Alltoall, with 16 processes, we recommend running the osu_alltoall benchmark in the following manner:

$ mpirun_rsh -np 16 -hostfile mf ./osu_alltoall

Where hostfile “mf” has the following contains:

node1
node1
node1
node1
node1
node1
node1
node1
node1
node2
node2
node2
node2
node2
node2
node2
node2

7.2.4 Running Benchmarks with CUDA/OpenACC Extensions

The following benchmarks have been extended to evaluate performance of MPI communication from and to buffers on NVIDIA GPU devices.

osu_bibw Bidirectional Bandwidth Test
osu_bw Bandwidth Test
osu_latencyLatency Test
osu_alltoallMPI_Alltoall Latency Test
osu_gather MPI_Gather Latency Test
osu_scatter MPI_Scatter Latency Test

Each of the pt2pt benchmarks takes two input parameters. The first parameter indicates the location of the buffers at rank 0 and the second parameter indicates the location of the buffers at rank 1. The value of each of these parameters can be either ’H’ or ’D’ to indicate if the buffers are to be on the host or on the device, respectively. When no parameters are specified, the buffers are allocated on the host.

The collective benchmarks will use buffers allocated on the device if the -d option is used otherwise the buffers will be allocated on the host.

If both CUDA and OpenACC support is enabled, you can switch between using CUDA and OpenACC to allocate your device buffers by specifying the ’-d cuda’ or ’-d openacc’ option to the benchmark. Please use the ’-h’ option for more info.

GPU affinity for processes should be set before MPI_Init is called. The process rank on a node is normally used to do this and different MPI launchers expose this information through different environment variables. The benchmarks use an environment variable called LOCAL_RANK to get this information. The script like get_local_rank provided alongside the benchmarks can be used to export this environment variable when using mpirun_rsh. This can be adapted to work with other MPI launchers and libraries.

Examples:

$ mpirun_rsh -np 2 host0 host0 MV2_USE_CUDA=1 get_local_rank ./osu_latency D D

In this run, assuming host0 has two GPU devices, the latency test allocates buffers on GPU device 0 at rank 0 and on GPU device 1 at rank 1.

$ mpirun_rsh -np 2 host0 host1 MV2_USE_CUDA=1 get_local_rank ./osu_bw D H

In this run, the bandwidth test allocates buffers on the GPU device at rank 0 (host0) and on the host at rank 1 (host1).

8 Scalability features and Performance Tuning for Large Scale Clusters

MVAPICH2 provides many different parameters for tuning performance for a wide variety of platforms and applications. This section deals with tuning CH3-based interfaces. These parameters can be either compile time parameters or runtime parameters. Please refer to Section 8 for a complete description of all these parameters. In this section we classify these parameters depending on what you are tuning for and provide guidelines on how to use them.

8.1 Optimizations for homogeneous clusters

MVAPICH2 internally detects the heterogeneity of the cluster in terms of processor and network interface type. Set parameter MV2_HOMOGENEOUS_CLUSTER to 1 to skip this detection, if the user already knows that cluster is homogeneous.

8.2 Job Launch Tuning

Starting with version 1.2, MVAPICH2 has a new, scalable job launcher – mpirun_rsh which uses a tree based mechanism to spawn processes. The degree of this tree is determined dynamically to keep the depth low. For large clusters, it might be beneficial to further flatten the tree by specifying a higher degree. The degree can be overridden with the environment variable MV2_MT_DEGREE (see 13.8).

MVAPICH2 can use a faster, hierarchical launching mechanism on large clusters. This is enabled manually using MV2_FASTSSH_THRESHOLD 13.5).

When the number of nodes involved is beyond 8k, the mpirun_rsh uses a file-based communication scheme to create the hierarchical tree. The default value can be overridden with the environment variable MV2_NPROCS_THRESHOLD (see 13.6).

8.3 Basic QP Resource Tuning

The following parameters affect memory requirements for each QP.

MV2_DEFAULT_MAX_SEND_WQE and MV2_DEFAULT_MAX_RECV_WQE control the maximum number of WQEs per QP and MV2_MAX_INLINE_SIZE controls the maximum inline size. Reducing the values of these two parameters leads to less memory consumption. They are especially important for large scale clusters with a large amount of connections and multiple rails.

These two parameters are run-time adjustable. Please refer to Sections 11.18 and 11.32 for details.

8.4 RDMA Based Point-to-Point Tuning

The following parameters are important in tuning the memory requirements for adaptive rdma fast path feature.

MV2_RDMA_FAST_PATH_BUF_SIZE is the size of each buffer used in RDMA fast path communication.

MV2_NUM_RDMA_BUFFER is number of buffers used for the RDMA fast path communication.

On the other hand, the product of MV2_RDMA_FAST_PATH_BUF_SIZE and
MV2_NUM_RDMA_BUFFER generally is a measure of the amount of memory registered for eager message passing. These buffers are not shared across connections.

8.5 Shared Receive Queue (SRQ) Tuning

The main environmental parameters controlling the behavior of the Shared Receive Queue design are:

MV2_SRQ_MAX_SIZE is the maximum size of the Shared Receive Queue (default 4096). You may increase this to value 8192 if the application requires very large number of processors. The application will start by only using MV2_SRQ_SIZE buffers (default 256) and will double this value on every SRQ limit event(upto MV2_SRQ_MAX_SIZE). For long running applications this re-size should show little effect. If needed, the MV2_SRQ_SIZE can be increased to 1024 or higher as needed for applications.
MV2_SRQ_LIMIT defines the low water-mark for the flow control handler. This can be reduced if your aim is to reduce the number of interrupts.

8.6 eXtended Reliable Connection (XRC)

MVAPICH2 now supports the eXtended Reliable Connection (XRC) transport available in recent Mellanox HCAs. This transport helps reduce the number of QPs needed on multi-core systems. Set MV2_USE_XRC (11.106) to use XRC with MVAPICH2.

8.7 Shared Memory Tuning

MVAPICH2 uses shared memory communication channel to achieve high-performance message passing among processes that are on the same physical node. The two main parameters which are used for tuning shared memory performance for small messages are SMPI_LENGTH_QUEUE (Section 11.111), and SMP_EAGER_SIZE (Section 11.110). The two main parameters which are used for tuning shared memory performance for large messages are SMP_SEND_BUF_SIZE (Section 11.113) and
SMP_NUM_SEND_BUFFER (Section 11.112).

SMPI_LENGTH_QUEUE is the size of the shared memory buffer which is used to store outstanding small and control messages. SMP_EAGER_SIZE defines the switch point from Eager protocol to Rendezvous protocol.

Messages larger than SMP_EAGER_SIZE are packetized and sent out in a pipelined manner.
SMP_SEND_BUF_SIZE is the packet size, i.e. the send buffer size. SMP_NUM_SEND_BUFFER is the number of send buffers.

8.8 On-demand Connection Management Tuning

MVAPICH2 uses on-demand connection management to reduce the memory usage of MPI library. There are 4 parameters to tune connection manager: MV2_ON_DEMAND_THRESHOLD (Section 11.44),
MV2_CM_RECV_BUFFERS (Section 11.10), MV2_CM_TIMEOUT (Section 11.12), and
MV2_CM_SPIN_COUNT (Section 11.11). The first one applies to OFA-IB-CH3, OFA-iWARP-CH3 and uDAPL-CH3 interfaces and the other three only apply to OFA-IB-CH3 interface.

MV2_ON_DEMAND_THRESHOLD defines threshold for enabling on-demand connection management scheme. When the size of the job is larger than the threshold value, on-demand connection management will be used.

MV2_CM_RECV_BUFFERS defines the number of buffers used by connection manager to establish new connections. These buffers are quite small and are shared for all connections, so this value may be increased to 8192 for large clusters to avoid retries in case of packet drops.

MV2_CM_TIMEOUT is the timeout value associated with connection management messages via UD channel. Decreasing this value may lead to faster retries but at the cost of generating duplicate messages.

MV2_CM_SPIN_COUNT is the number of the connection manager polls for new control messages from UD channel for each interrupt. This may be increased to reduce the interrupt overhead when many incoming control messages from UD channel at the same time.

8.9 Scalable Collectives Tuning

MVAPICH2 uses shared memory to optimize the performance for many collective operations: MPI_Allreduce, MPI_Reduce, MPI_Barrier, and MPI_Bcast.

We use shared-memory based collective for most small and medium sized messages and fall back to the default point-to-point based algorithms for very large messages. The upper-limits for shared-memory based collectives are tunable parameters that are specific to each collective operation. We have variables such as MV2_SHMEM_ALLREDUCE_MSG (11.61), MV2_SHMEM_REDUCE_MSG (11.67) and
MV2_SHMEM_BCAST_MSG (11.63), for MPI_Allreduce, MPI_Reduce and MPI_Bcast collective operations. The default values for these variables have been set through experimental analysis on some of our clusters and a few large scale clusters, such as the TACC Ranger. Users can choose to set these variables at job-launch time to tune the collective algorithms on different platforms.

8.9.1 Optimizations for MPI_Bcast

MVAPICH2 supports a 2-level point-to-point tree-based “Knomial” algorithm for small messages for the MPI_Bcast operation. MVAPICH2 also offers improved designs that deliver better performance for medium and large message lengths.

8.9.2 Optimizations for MPI_Reduce and MPI_Allreduce

In this release, we have introduced new 2-level algorithms for MPI_Reduce and MPI_Allreduce operations, along the same lines as MPI_Bcast. Pure Shared-memory based algorithms cannot be used for larger messages for reduction operations, because the node-leader processes become the bottleneck, as they have to perform the reduction operation on the entire data block. We now rely on shared-memory algorithms for MPI_Reduce and MPI_Allreduce for message sizes set by the thresholds MV2_SHMEM_ALLREDUCE_MSG (11.61), MV2_SHMEM_REDUCE_MSG (11.67), the new 2-level algorithms for medium sized messages and the default point-to-point based algorithms for large messages. We have introduced two new run-time variables MV2_ALLREDUCE_2LEVEL_MSG (11.1) and MV2_REDUCE_2LEVEL_MSG (11.57) to determine when to fall back to the default point-to-point based algorithms.

8.9.3 Optimizations for MPI_Gather and MPI_Scatter

MVAPICH2 supports two new optimized algorithms for MPI_Gather and MPI_Scatter operations – the “Direct” and the multi-core aware “2-level” algorithms. Both these algorithms perform significantly better than the default binomial-tree pattern. The “Direct” algorithm is however inherently not very scalable and can be used when the communicator size is less than 1K processes. We switch over to the 2-level algorithms for larger system sizes. For MPI_Gather, we use different algorithms depending of the system size. For small system sizes (up to 386 cores), we use the “2-level” algorithm following by the “Direct” algorithm. For medium system sizes (up to 1k), we use “Binomial” algorithm following by the “Direct” algorithm. It’s possible to set the switching point between algorithms using the run-time parameter MV2_GATHER_SWITCH_PT (11.101). For MPI_Scatter, when the system size is lower than 512 cores, we use the “Binomial” algorithm for small message sizes following by the “2-level” algorithm for medium message sizes and the “Direct” algorithm for large message sizes. Users can define the threshold for small and medium message sizes using the run-time parameters MV2_SCATTER_SMALL_MSG (11.102) and MV2_SCATTER_MEDIUM_MSG (11.103). Users can also choose to use only one of these algorithms by toggling the run-time parameters  11.80 and  11.104 for MPI_Gather and  11.81 and  11.105 for MPI_Scatter.

8.10 Process Placement on Multi-core platforms

Process placement has a significant impact on performance of applications. Depending on your application communication patterns, various process placements can result in performance gains. In Section 6.5, we have described the usage of “bunch” and “scatter” placement modes provided by MVAPICH2. Using these modes, one can control the placement of processes within a particular node. Placement of processes across nodes can be controlled by adjusting the order of MPI ranks. For example, the following command launches jobs in block fashion.

$ mpirun_rsh -np 4 n0 n0 n1 n1 MV2_CPU_BINDING_POLICY=bunch ./a.out

The following command launches jobs in a cyclic fashion.

$ mpirun_rsh -np 4 n0 n1 n0 n1 MV2_CPU_BINDING_POLICY=scatter ./a.out

We have noted that the HPL (High-Performance Linpack) benchmark performs better when using block distribution.

8.11 HugePage Support

MVAPICH2 uses HugePages(2MB) by default for communication buffers if they are configured on the system. The run-time variable, MV2_USE_HUGEPAGES( 11.114) can be used to control the behavior of this feature.

In order to use HugePages, Make sure HugePages are configured on all nodes. The number of HugePages can be configured by setting vm.nr_hugepages kernel parameter to a suitable value. For example, to allocate a 1GB HugePage pool, execute(as root):

$ echo 512 > /proc/sys/vm/nr_hugepages
or
$ sysctl -w vm.nr_hugepages = 512

9 FAQ and Troubleshooting with MVAPICH2

Based on our experience and feedback we have received from our users, here we include some of the problems a user may experience and the steps to resolve them. If you are experiencing any other problem, please feel free to contact us by sending an email to mvapich-discuss@cse.ohio-state.edu.

MVAPICH2 can be used over eight underlying interfaces, namely OFA-IB-CH3, OFA-IB-Nemesis, OFA-IWARP-CH3, OFA-RoCE-CH3, PSM-CH3, uDAPL-CH3, TCP/IP-CH3 and TCP/IP-Nemesis. Based on the underlying library being utilized, the troubleshooting steps may be different. We have divided the troubleshooting tips into four sections: General troubleshooting and Troubleshooting over any one of the five transport interfaces.

9.1 General Questions and Troubleshooting

9.1.1 MVAPICH2 is failed to register memory with Infiniband HCA

OFED provides network vendor specific kernel module paramters to control the size of Memory translation table(MTT) used to map virtual to physical address. This will limit the amount of physical memory can be registered with infiniband device. The following two parameters are provided to control the size of this table.

  1. log_num_mtt
  2. log_mtts_per_seg

The amount of memory that can be registered is calculated by

max_reg_mem = (2log_num_mtt) * (2log_mtts_per_seg) * PAGE_SIZE

It is recommended to adjust log_num_mtt to allow at least twice the amount of physical memory on your machine. For example, if a node has 64 GB of memory and a 4 KB page size, log_num_mtt should be set to 24 and (assuming log_mtts_per_seg is set to 1)

These parameters are set on the mlx4_core module in /etc/modprobe.conf
options mlx4_core log_num_mtt=24

9.1.2 Invalid Communicators Error

This is a problem which typically occurs due to the presence of multiple installations of MVAPICH2 on the same set of nodes. The problem is due to the presence of mpi.h other than the one, which is used for executing the program. This problem can be resolved by making sure that the mpi.h from other installation is not included.

9.1.3 Are fork() and system() supported?

fork() and system() is supported for the OpenFabrics device as long as the kernel is being used is Linux 2.6.16 or newer. Additionally, the version of OFED used should be 1.2 or higher. The environment variable IBV_FORK_SAFE=1 must also be set to enable fork support.

9.1.4 MPI+OpenMP shows bad performance

MVAPICH2 uses CPU affinity to have better performance for single-threaded programs. For multi-threaded programs, e.g. MPI+OpenMP, it may schedule all the threads of a process to run on the same CPU. CPU affinity should be disabled in this case to solve the problem, i.e. set MV2_ENABLE_AFFINITY to 0. In addition, please read Section  6.17 on using MVAPICH2 in multi-threaded environments. We also recommend using the compiler/platform specific run-time options to bind the OpenMP threads to processors. Please refer to Section ( 6.18) for more information.

9.1.5 Error message “No such file or directory” when using Lustre file system

If you are using ADIO support for Lustre, please make sure of the following:
– Check your Lustre setup
– You are able to create, read to and write from files in the Lustre mounted directory
– The directory is mounted on all nodes on which the job is executed
– The path to the file is correctly specified
– The permissions for the file or directory are correctly specified

9.1.6 Program segfaults with “File locking failed in ADIOI_Set_lock”

If you are using ADIO support for Lustre, the recent Lustre releases require an additional mount option to have correct file locks. Please include the following option with your lustre mount command: “-o localflock”.

$ mount -o localflock -t lustre xxxx@o2ib:/datafs /mnt/datafs

9.1.7 Running MPI programs built with gfortran

MPI programs built with gfortran might not appear to run correctly due to the default output buffering used by gfortran. If it seems there is an issue with program output, the GFORTRAN_UNBUFFERED_ALL variable can be set to “y” and exported into the environment before using the mpiexec or mpirun_rsh command to launch the program, as below:

$ export GFORTRAN_UNBUFFERED_ALL=y

Or, if using mpirun_rsh, export the environment variable as in the example:

$ mpirun_rsh -np 2 n1 n2 GFORTRAN_UNBUFFERED_ALL=y ./a.out

9.1.8 How do I obtain MVAPICH2 version and configuration information?

The mpiname application is provided with MVAPICH2 to assist with determining the MPI library version and related information. The usage of mpiname is as follows:

$ mpiname [OPTION]

Print MPI library information. With no OPTION, the output is the same as -v.

-a print all information

-c print compilers

-d print device

-h display this help and exit

-n print the MPI name

-o print configuration options

-r print release date

-v print library version

9.1.9 How do I compile my MPI application with static libraries, and not use shared libraries?

Even if MVAPICH2 installed on your cluster was configured to use shared libraries, it is possible to use static libraries. Use the command below to use static libraries:

$ mpicc -noshlib -o cpi cpi.c

9.1.10 Does MVAPICH2 work across AMD and Intel systems?

Yes, as long as you compile MVAPICH2 and your programs on one of the systems, either AMD or Intel, and run the same binary across the systems. MVAPICH2 has platform specific parameters for performance optimizations and it may not work if you compile MVAPICH2 and your programs on different systems and try to run the binaries together.

9.1.11 I want to enable debugging for my build. How do I do this?

We recommend that you enable debugging when you intend to take a look at back traces of processes in GDB (or other debuggers). You can use the following configure options to enable debugging: --enable-g=dbg --disable-fast.

Additionally:

9.1.12 How can I run my application with a different group ID?

You can specify a different group id for your MPI application using the -sg group option to mpirun_rsh. The following example executes a.out on host1 and host2 using secondarygroup as their group id.

$ mpirun_rsh -sg secondarygroup -np 2 host1 host2 ./a.out

9.2 Issues and Failures with Job launchers

9.2.1 Can I use MVAPICH2 on PBS/Torque Clusters

You can use MVAPICH2 for clusters administered by PBS/Torque. If you are a cluster user (not an administrator), please ask your cluster administrator to install the OSC mpiexec. If you are a cluster administrator, please follow the instructions below.

You will need to download mpiexec from Ohio Supercomputer enter (OSC) at the following link.

Please note that this mpiexec is different from the mpiexec provided within MPICH2. Also note that you do not need to use either mpirun_rsh or mpiexec.hydra on a cluster that is administered with PBS/Torque. You may also choose to remove the MVAPICH2 mpiexecs from the install path to shield your users from making a mistake of trying to run MPI jobs with a wrong launcher.

Usage of OSC mpiexec is simple. An example of using this launcher is as follows:

$ /path/to/osc/mpiexec -np 4 ./a.out

9.2.2 /usr/bin/env: mpispawn: No such file or directory

If mpirun_rsh fails with this error message, it was unable to locate a necessary utility. This can be fixed by ensuring that all MVAPICH2 executables are in the PATH on all nodes. If PATHs cannot be setup as mentioned, then invoke mpirun_rsh with a path prefix. For example:

$ /path/to/mpirun_rsh -np 2 node1 node2 ./mpi_proc

9.2.3 TotalView complains that “The MPI library contains no suitable type definition for struct MPIR_PROCDESC”

Ensure that the MVAPICH2 job launcher mpirun_rsh is compiled with debug symbols. Details are available in Section 5.2.11.

9.3 Problems Building MVAPICH2

9.3.1 Unable to convert MPI_SIZEOF_AINT to a hex string

configure: error: Unable to convert MPI_SIZEOF_AINT to a hex string.
This is either because we are building on a very strange platform or there is a bug somewhere in configure.

This error can be misleading. The problem is often not that you’re building on a strange platform, but that there was some problem running an executable that made configure have trouble determining the size of a datatype. The true problem is often that you’re trying to link against a library that is not found in your system’s default path for linking at runtime. Please check that you’ve properly set LD_LIBRARY_PATH or used the correct rpath settings in LDFLAGS.

9.3.2 Cannot Build with the PathScale Compiler

There is a known bug with the PathScale compiler (before version 2.5) when building MVAPICH2. This problem will be solved in the next major release of the PathScale compiler. To work around this bug, use the the “-LNO:simd=0” C compiler option. This can be set in the build script similarly to:

export CC="pathcc -LNO:simd=0"

Please note the use of double quotes. If you are building shared libraries and are using the PathScale compiler (version below 2.5), then you should add “-g” to your CFLAGS, in order to get around a compiler bug.

9.3.3 Libtool has a problem linking with non-GNU compiler (like PGI)

If you are using a compiler that is not recognized by autoconf as a GNU compiler, Libtool uses an default library search path to look for shared objects which is ”/lib /usr/lib /usr/local/lib”. Then, if your libraries are not in one of these paths, MVAPICH2 may fail to link properly.

You can work around this issue by adding the following configure flags:

./configure \  
lt_cv_sys_lib_search_path_spec="/lib64 /usr/lib64 /usr/local/lib64"   \  
lt_cv_sys_lib_dlsearch_path_spec="/lib64 /usr/lib64 /usr/local/lib64" \  
... ...

The above example considers that the correct library search path for your system is ”/lib64 /usr/lib64 /usr/local/lib64”.

9.4 With OFA-IB-CH3 Interface

9.4.1 Cannot Open HCA

The above error reports that the InfiniBand Adapter is not ready for communication. Make sure that the drivers are up. This can be done by executing the following command which gives the path at which drivers are setup.

$ locate libibverbs

9.4.2 Checking state of IB Link

In order to check the status of the IB link, one of the OFED utilities can be used: ibstatus, ibv_devinfo.

9.4.3 Creation of CQ or QP failure

A possible reason could be inability to pin the memory required. Make sure the following steps are taken.

  1. In /etc/security/limits.conf add the following
    * soft memlock phys_mem_in_KB

  2. After this, add the following to /etc/init.d/sshd
    ulimit -l  phys_mem_in_KB

  3. Restart sshd

With some distros, we’ve found that adding the ulimit -l line to the sshd init script is no longer necessary. For instance, the following steps work for our RHEL5 systems.

  1. Add the following lines to /etc/security/limits.conf
    * soft memlock unlimited  
    * hard memlock unlimited

  2. Restart sshd

9.4.4 Hang with the HSAM Functionality

HSAM functionality uses multi-pathing mechanism with LMC functionality. However, some versions of OpenFabrics Drivers (including OpenFabrics Enterprise Distribution (OFED) 1.1) and using the Up*/Down* routing engine do not configure the routes correctly using the LMC mechanism. We strongly suggest to upgrade to OFED 1.2, which supports Up*/Down* routing engine and LMC mechanism correctly.

9.4.5 Failure with Automatic Path Migration

MVAPICH2 (OFA-IB-CH3) provides network fault tolerance with Automatic Path Migration (APM). However, APM is supported only with OFED 1.2 onwards. With OFED 1.1 and prior versions of OpenFabrics drivers, APM functionality is not completely supported. Please refer to Section 11.76 and section 11.77

9.4.6 Error opening file

If you configure MVAPICH2 with RDMA_CM and see this error, you need to verify if you have setup up the local IP address to be used by RDMA_CM in the file /etc/mv2.conf. Further, you need to make sure that this file has the appropriate file read permissions. Please follow Section 6.16 for more details on this.

9.4.7 RDMA CM Address error

If you get this error, please verify that the IP address specified /etc/mv2.conf is correctly specified with the IP address of the device you plan to use RDMA_CM with.

9.4.8 RDMA CM Route error

If see this error, you need to check whether the specified network is working or not.

9.5 With OFA-iWARP-CH3 Interface

9.5.1 Error opening file

If you configure MVAPICH2 with RDMA_CM and see this error, you need to verify if you have setup up the local IP address to be used by RDMA_CM in the file /etc/mv2.conf. Further, you need to make sure that this file has the appropriate file read permissions. Please follow Section 5.2.6 for more details on this.

9.5.2 RDMA CM Address error

If you get this error, please verify that the IP address specified /etc/mv2.conf is correctly specified with the IP address of the device you plan to use RDMA_CM with.

9.5.3 RDMA CM Route error

If see this error, you need to check whether the specified network is working or not.

9.6 With uDAPL-CH3 Interface

9.6.1 Cannot Open IA

If you configure MVAPICH2 with uDAPL and see this error, you need to check whether you have specified the correct uDAPL service provider (Section 5.2.8). If you have specified the uDAPL provider but still see this error, you need to check whether the specified network is working or not. If you are using OpenFabrics software stack on Linux, the default DAPL provider is OpenIB-cma for DAPL-1.2, and ofa-v2-ib0 for DAPL-2.0. If you are using Solaris, the default DAPL provider is ibd0.

9.6.2 DAT Insufficient Resource

If you configure MVAPICH2 with uDAPL and see this error, you need to reduce the value of the environmental variables RDMA_DEFAULT_MAX_SEND_WQE and/or RDMA_DEFAULT_MAX_RECV_WQE depending on the underlying network.

9.6.3 Cannot Find libdat.so

If you get the error: “error while loading shared libraries, libdat.so”, The location of the dat shared library is incorrect. You need to find the correct path of libdat.so and export LD_LIBRARY_PATH to this correct location. For example:

$ mpirun_rsh -np 2 n1 n2 LD_LIBRARY_PATH=/path/to/libdat.so ./a.out

9.6.4 uDAPL over IB Does Not Scale Beyond 256 Nodes with RDMA_CM Provider

We recommend that uDAPL IB consumers needing large scale-out use socket cm provider (libdaplscm.so) in lieu of rdma_cm (libdaplcma.so). iWARP users can remain using uDAPL rdma_cm provider. For detailed discussion of this issue please refer to:

http://lists.openfabrics.org/pipermail/general/2008-June/051814.html

9.7 Checkpoint/Restart

9.7.1 Failure during Restart

Please make sure the following things for a successful restart:

The following things can cause a restart to fail:

FAQ regarding Berkeley Lab Checkpoint/Restart (BLCR) can be found at:
 http://upc-bugs.lbl.gov/blcr/doc/html/FAQ.html And the user guide for BLCR can be found at  http://upc-bugs.lbl.gov/blcr/doc/html/BLCR_Users_Guide.html

If you encounter any problem with the Checkpoint/Restart support, please feel free to contact us at mvapich-discuss@cse.ohio-state.edu.

10 MVAPICH2 General Parameters

10.1 MV2_IGNORE_SYSTEM_CONFIG

If set, the system configuration file is not processed.

10.2 MV2_IGNORE_USER_CONFIG

If set, the user configuration file is not processed.

10.3 MV2_USER_CONFIG

Specify the path of a user configuration file for mvapich2. If this is not set the default path of “ /.mvapich2.conf” is used.

10.4 MV2_DEBUG_CORESIZE

Set the limit for the core size resource. It allows to specify the maximum size for a core dump to be generated. It only set the soft limit and it has the respect the hard value set on the nodes.

It is similar to the ulimit -c <coresize> that can be run in the shell, but this will only apply to the MVAPICH2 processes (MPI processes, mpirun_rsh, mpispawn).

Examples:

10.5 MV2_DEBUG_SHOW_BACKTRACE

Show a backtrace when a process fails on errors like ”Segmentation faults”, ”Bus error”, ”Illegal Instruction”, ”Abort” or ”Floating point exception”.

If your application uses the static version of the MVAPICH2 library, you have to link your application with the -rdynamic flag in order to see the function names in the backtrace. For more information, see the backtrace manpage.

11 MVAPICH2 Parameters (CH3-Based Interfaces)

11.1 MV2_ALLREDUCE_2LEVEL_MSG

This parameter can be used to determine the threshold for the 2-level Allreduce algorithm. We now use the shared-memory-based algorithm for messages smaller than the MV2_SHMEM_ALLREDUCE_MSG threshold (11.61), the 2-level algorithm for medium sized messages up to the threshold defined by this parameter. We use the default point-to-point algorithms messages larger than this threshold.

11.2 MV2_CKPT_AGGREGATION_BUFPOOL_SIZE

This parameter determines the size of the buffer pool reserved for use in checkpoint aggregation. Note that this variable can be set with suffixes such as ‘K’/‘k’, ‘M’/‘m’ or ‘G’/‘g’ to denote Kilobyte, Megabyte or Gigabyte respectively.

11.3 MV2_CKPT_AGGREGATION_CHUNK_SIZE

The checkpoint data that has been coalesced into the buffer pool, is written to the back-end file system, with the value of this parameter as the chunk size. Note that this variable can be set with suffixes such as ‘K’/‘k’, ‘M’/‘m’ or ‘G’/‘g’ to denote Kilobyte, Megabyte or Gigabyte respectively.

11.4 MV2_CKPT_FILE

This parameter specifies the path and the base file name for checkpoint files of MPI processes. The checkpoint files will be named as $MV2_CKPT_FILE.<number of checkpoint>.<process rank>, for example, /tmp/ckpt.1.0 is the checkpoint file for process 0’s first checkpoint. To checkpoint on network-based file systems, user just need to specify the path to it, such as /mnt/pvfs2/my_ckpt_file.

11.5 MV2_CKPT_INTERVAL

This parameter can be used to enable automatic checkpointing. To let MPI job console automatically take checkpoints, this value needs to be set to the desired checkpointing interval. A zero will disable automatic checkpointing. Using automatic checkpointing, the checkpoint file for the MPI job console will be named as $MV2_CKPT_FILE.<number of checkpoint>.auto. Users need to use this file for restart.

11.6 MV2_CKPT_MAX_SAVE_CKPTS

This parameter is used to limit the number of checkpoints saved on file system to save the file system space. When set to a positive value N, only the last N checkpoints will be saved.

11.7 MV2_CKPT_NO_SYNC

When this parameter is set to any value, the checkpoints will not be required to sync to disk. It can reduce the checkpointing delay in many cases. But if users are using local file system, or any parallel file system with local cache, to store the checkpoints, it is recommended not to set this parameter because otherwise the checkpoint files will be cached in local memory and will likely be lost upon failure.

11.8 MV2_CKPT_USE_AGGREGATION

This parameter enables/disables Checkpoint aggregation scheme at run time. It is set to ’1’(enabled) by default, when the user enables Checkpoint/Restart functionality at configure time, or when the user explicitly configures MVAPICH2 with aggregation support. Please note that, to use aggregation support, each node needs to be properly configured with FUSE library (cf section 6.15.1).

11.9 MV2_DEBUG_FT_VERBOSE

This parameter enables/disables the debug output for Fault Tolerance features (Checkpoint/Restart and Migration).

Note: All debug output is disabled when MVAPICH2 is configured with the --enable-fast=ndebug option.

11.10 MV2_CM_RECV_BUFFERS

This defines the number of buffers used by connection manager to establish new connections. These buffers are quite small and are shared for all connections, so this value may be increased to 8192 for large clusters to avoid retries in case of packet drops.

11.11 MV2_CM_SPIN_COUNT

This is the number of the connection manager polls for new control messages from UD channel for each interrupt. This may be increased to reduce the interrupt overhead when many incoming control messages from UD channel at the same time.

11.12 MV2_CM_TIMEOUT

This is the timeout value associated with connection management messages via UD channel. Decreasing this value may lead to faster retries but at the cost of generating duplicate messages.

11.13 MV2_CPU_MAPPING

This allows users to specify process to CPU (core) mapping. The detailed usage of this parameter is described in Section 6.5.2. This parameter will not take effect if either MV2_ENABLE_AFFINITY or MV2_USE_SHARED_MEM run-time parameters are set to 0, or if the library was configured with the “–disable-hwloc” option. MV2_CPU_MAPPING is currently not supported on Solaris.

11.14 MV2_CPU_BINDING_POLICY

This allows users to specify process to CPU (core) mapping with the CPU binding policy. The detailed usage of this parameter is described in Section 6.5.1. This parameter will not take effect: if
MV2_ENABLE_AFFINITY or MV2_USE_SHARED_MEM run-time parameters are set to 0; or
MV2_ENABLE_AFFINITY is set to 1 and MV2_CPU_MAPPING is set, or if the library was configured with the “–disable-hwloc” option. The value of MV2_CPU_BINDING_POLICY can be “bunch” or “scatter”. When this parameter takes effect and its value isn’t set, “bunch” will be used as the default policy.

11.15 MV2_CPU_BINDING_LEVEL

This allows users to specify process to CPU (core) mapping at different binding level. The detailed usage of this parameter is described in Section 6.5.1. This parameter will not take effect: if MV2_ENABLE_AFFINITY or MV2_USE_SHARED_MEM run-time parameters are set to 0; or MV2_ENABLE_AFFINITY is set to 1 and MV2_CPU_MAPPING is set, or if the library was configured with the “–disable-hwloc” option. The value of MV2_CPU_BINDING_LEVEL can be “core”, “socket”, or “numanode”. When this parameter takes effect and its value isn’t set, “core” will be used as the default binding level.

11.16 MV2_SHOW_CPU_BINDING

This shows the current process to CPU (core) bindings of all processes on a node which contains the mpi rank 0 process.

11.17 MV2_DAPL_PROVIDER

This is to specify the underlying uDAPL-CH3 library that the user would like to use if MVAPICH2 is built with uDAPL-CH3.

11.18 MV2_DEFAULT_MAX_SEND_WQE

This specifies the maximum number of send WQEs on each QP. Please note that for OFA-IB-CH3 and OFA-iWARP-CH3, the default value of this parameter will be 16 if the number of processes is larger than 256 for better memory scalability.

11.19 MV2_DEFAULT_MAX_RECV_WQE

This specifies the maximum number of receive WQEs on each QP (maximum number of receives that can be posted on a single QP).

11.20 MV2_DEFAULT_MTU

The internal MTU size. For OFA-IB-CH3, this parameter should be a string instead of an integer. Valid values are: IBV_MTU_256, IBV_MTU_512, IBV_MTU_1024, IBV_MTU_2048, IBV_MTU_4096.

11.21 MV2_DEFAULT_PKEY

Select the partition to be used for the job.

11.22 MV2_ENABLE_AFFINITY

Enable CPU affinity by setting MV2_ENABLE_AFFINITY to 1 or disable it by setting
MV2_ENABLE_AFFINITY to 0. MV2_ENABLE_AFFINITY is currently not supported on Solaris. CPU affinity is also not supported if MV2_USE_SHARED_MEM is set to 0.

11.23 MV2_GET_FALLBACK_THRESHOLD

This defines the threshold beyond which the MPI_Get implementation is based on direct one sided RDMA operations.

11.24 MV2_IBA_EAGER_THRESHOLD

This specifies the switch point between eager and rendezvous protocol in MVAPICH2. For better performance, the value of MV2_IBA_EAGER_THRESHOLD should be set the same as MV2_VBUF_TOTAL_SIZE.

11.25 MV2_IBA_HCA

This specifies the HCA’s to be used for performing network operations.

11.26 MV2_INITIAL_PREPOST_DEPTH

This defines the initial number of pre-posted receive buffers for each connection. If communication happen for a particular connection, the number of buffers will be increased to
RDMA_PREPOST_DEPTH.

11.27 MV2_IWARP_MULTIPLE_CQ_THRESHOLD

This defines the process size beyond which we use multiple completion queues for iWARP interface.

11.28 MV2_KNOMIAL_INTRA_NODE_FACTOR

This defines the degree of the knomial operation during the intra-node knomial broadcast phase.

11.29 MV2_KNOMIAL_INTER_NODE_FACTOR

This defines the degree of the knomial operation during the inter-node knomial broadcast phase.

11.30 MV2_LIMIC_GET_THRESHOLD

This defines the message size threshold (in bytes) beyond which intra-node MPI_Get operations are optimized using the LiMIC2 Kernel module.

11.31 MV2_LIMIC_PUT_THRESHOLD

This defines the message size threshold (in bytes) beyond which intra-node MPI_Put operations are optimized using the LiMIC2 Kernel module.

11.32 MV2_MAX_INLINE_SIZE

This defines the maximum inline size for data transfer. Please note that the default value of this parameter will be 0 when the number of processes is larger than 256 to improve memory usage scalability.

11.33 MV2_MAX_NUM_WIN

Maximum number of RMA windows that can be created and active concurrently. Typically this value is sufficient for most applications. Increase this value to the number of windows your application uses

11.34 MV2_NDREG_ENTRIES

This defines the total number of buffers that can be stored in the registration cache. It has no effect if MV2_USE_LAZY_MEM_UNREGISTER is not set. A larger value will lead to less frequent lazy de-registration.

11.35 MV2_NUM_HCAS

This parameter indicates number of InfiniBand adapters to be used for communication on an end node.

11.36 MV2_NUM_PORTS

This parameter indicates number of ports per InfiniBand adapter to be used for communication per adapter on an end node.

11.37 MV2_NUM_SA_QUERY_RETRIES

Number of times the MPI library will attempt to query the subnet to obtain the path record information before giving up.

11.38 MV2_NUM_QP_PER_PORT

This parameter indicates number of queue pairs per port to be used for communication on an end node. This is useful in the presence of multiple send/recv engines available per port for data transfer.

11.39 MV2_RAIL_SHARING_POLICY

This specifies the policy that will be used to assign HCAs to each of the processes. In the previous versions of MVAPICH2 it was known as MV2_SM_SCHEDULING.

11.40 MV2_RAIL_SHARING_LARGE_MSG_THRESHOLD

This specifies the threshold for the message size beyond which striping will take place. In the previous versions of MVAPICH2 it was known as MV2_STRIPING_THRESHOLD

11.41 MV2_PROCESS_TO_RAIL_MAPPING

When MV2_RAIL_SHARING_POLICY is set to the value “FIXED_MAPPING” this variable decides the manner in which the HCAs will be mapped to the rails. The <CUSTOM LIST> is colon(:) separated list with the HCA ranks specified. e.g. 0:1:1:0. This list must map equally to the number of local processes on the nodes failing which, the default policy will be used. Similarly the number of processes on each node must be the same.

11.42 MV2_RDMA_FAST_PATH_BUF_SIZE

The size of the buffer used in RDMA fast path communication. This value will be ineffective if
MV2_USE_RDMA_FAST_PATH is not set

11.43 MV2_NUM_RDMA_BUFFER

The number of RDMA buffers used for the RDMA fast path. This fast path is used to reduce latency and overhead of small data and control messages. This value will be ineffective if MV2_USE_RDMA_FAST_PATH is not set.

11.44 MV2_ON_DEMAND_THRESHOLD

This defines threshold for enabling on-demand connection management scheme. When the size of the job is larger than the threshold value, on-demand connection management will be used.

11.45 MV2_HOMOGENEOUS_CLUSTER

Set this parameter to 1 on homogeneous clusters to optimize the job start-up

11.46 MV2_PREPOST_DEPTH

11.47 MV2_PREPOST_DEPTH

This defines the number of buffers pre-posted for each connection to handle send/receive operations.

11.48 MV2_PROCESS_TO_RAIL_MAPPING

This allows users to specify HCAs to be used by processes. The detailed usage of this parameter is described in Section 6.13.

11.49 MV2_PSM_DEBUG

This parameter enables the dumping of run-time debug counters from the MVAPICH2-PSM progress engine. Counters are dumped every PSM_DUMP_FREQUENCY seconds.

11.50 MV2_PSM_DUMP_FREQUENCY

This parameters sets the frequency for dumping MVAPICH2-PSM debug counters. Value takes effect only in PSM_DEBUG is enabled.

11.51 MV2_PUT_FALLBACK_THRESHOLD

This defines the threshold beyond which the MPI_Put implementation is based on direct one sided RDMA operations.

11.52 MV2_RAIL_SHARING_LARGE_MSG_THRESHOLD

This parameter specifies the message size above which we begin the stripe the message across multiple rails (if present).

11.53 MV2_RAIL_SHARING_POLICY

This parameter specifies the rail sharing policy for all message sizes if multiple rails are present.

11.54 MV2_RDMA_CM_ARP_TIMEOUT

This parameter specifies the ARP timeout to be used by RDMA CM module.

11.55 MV2_RDMA_CM_MAX_PORT

This parameter specifies the upper limit of the port range to be used by the RDMA CM module when choosing the port on which it listens for connections.

11.56 MV2_RDMA_CM_MIN_PORT

This parameter specifies the lower limit of the port range to be used by the RDMA CM module when choosing the port on which it listens for connections.

11.57 MV2_REDUCE_2LEVEL_MSG

This parameter can be used to determine the threshold for the 2-level reduce algorithm. We now use the shared-memory-based algorithm for messages smaller than the MV2_SHMEM_REDUCE_MSG (11.67), the 2-level algorithm for medium sized messages up to the threshold defined by this parameter. We use the default point-to-point algorithms messages larger than this threshold.

11.58 MV2_RNDV_PROTOCOL

The value of this variable can be set to choose different Rendezvous protocols. RPUT (default RDMA-Write) RGET (RDMA Read based), R3 (send/recv based).

11.59 MV2_R3_THRESHOLD

The value of this variable controls what message sizes go over the R3 rendezvous protocol. Messages above this message size use MV2_RNDV_PROTOCOL.

11.60 MV2_R3_NOCACHE_THRESHOLD

The value of this variable controls what message sizes go over the R3 rendezvous protocol when the registration cache is disabled (MV2_USE_LAZY_MEM_UNREGISTER=0). Messages above this message size use MV2_RNDV_PROTOCOL.

11.61 MV2_SHMEM_ALLREDUCE_MSG

The SHMEM AllReduce is used for messages less than this threshold.

11.62 MV2_SHMEM_BCAST_LEADERS

The number of leader processes that will take part in the SHMEM broadcast operation. Must be greater than the number of nodes in the job.

11.63 MV2_SHMEM_BCAST_MSG

The SHMEM bcast is used for messages less than this threshold.

11.64 MV2_SHMEM_COLL_MAX_MSG_SIZE

This parameter can be used to select the max buffer size of message for shared memory collectives.

11.65 MV2_SHMEM_COLL_NUM_COMM

This parameter can be used to select the number of communicators using shared memory collectives.

11.66 MV2_SHMEM_DIR

This parameter can be used to specify the path to the shared memory files for intra-node communication.

11.67 MV2_SHMEM_REDUCE_MSG

The SHMEM reduce is used for messages less than this threshold.

11.68 MV2_SM_SCHEDULING

11.69 MV2_SMP_USE_LIMIC2

This parameter enables/disables LiMIC2 at run time. It does not take effect if MVAPICH2 is not configured with –with-limic2.

11.70 MV2_SMP_USE_CMA

This parameter enables/disables CMA based intra-node communication at run time. It does not take effect if MVAPICH2 is not configured with –with-cma. When both –with-limic2 and –with-cma are included in the configure flags, this parameter is off by default. Please set MV2_SMP_USE_LIMIC2 to 0 in order to choose CMA if MVAPICH2 is configured with –with-limic2.

11.71 MV2_SRQ_LIMIT

This is the low water-mark limit for the Shared Receive Queue. If the number of available work entries on the SRQ drops below this limit, the flow control will be activated.

11.72 MV2_SRQ_MAX_SIZE

This is the maximum number of work requests allowed on the Shared Receive Queue. Upon receiving a SRQ limit event, the current value of MV2_SRQ_SIZE will be doubled or moved to the maximum of MV2_SRQ_MAX_SIZE, whichever is smaller

11.73 MV2_SRQ_SIZE

This is the initial number of work requests posted to the Shared Receive Queue.

11.74 MV2_STRIPING_THRESHOLD

This parameter specifies the message size above which we begin the stripe the message across multiple rails (if present).

11.75 MV2_SUPPORT_DPM

This option enables the dynamic process management interface and on-demand connection management.

11.76 MV2_USE_APM

This parameter is used for recovery from network faults using Automatic Path Migration. This functionality is beneficial in the presence of multiple paths in the network, which can be enabled by using lmc mechanism.

11.77 MV2_USE_APM_TEST

This parameter is used for testing the Automatic Path Migration functionality. It periodically moves the alternate path as the primary path of communication and re-loads another alternate path.

11.78 MV2_USE_BLOCKING

Setting this parameter enables MVAPICH2 to use blocking mode progress. MPI applications do not take up any CPU when they are waiting for incoming messages.

11.79 MV2_USE_COALESCE

Setting this parameter enables message coalescing to increase small message throughput

11.80 MV2_USE_DIRECT_GATHER

Use the “Direct” algorithm for the MPI_Gather operation. If this parameter is set to 0 at run-time, the “Direct” algorithm will not be invoked.

11.81 MV2_USE_DIRECT_SCATTER

Use the “Direct” algorithm for the MPI_Scatter operation. If this parameter is set to 0 at run-time, the “Direct” algorithm will not be invoked.

11.82 MV2_USE_HSAM

This parameter is used for utilizing hot-spot avoidance with InfiniBand clusters. To leverage this functionality, the subnet should be configured with lmc greater than zero. Please refer to section 6.19 for detailed information.

11.83 MV2_USE_IWARP_MODE

This parameter enables the library to run in iWARP mode. The library has to be built using the flag -DRDMA_CM for using this feature.

11.84 MV2_USE_LAZY_MEM_UNREGISTER

Setting this parameter enables MVAPICH2 to use memory registration cache.

11.85 MV2_USE_LIMIC_ONE_SIDED

Setting this parameter allows MVAPICH2 to use LiMIC2 kernel module for one sided communication within a node. Currently this optimization is available only with the Post-Wait/Start-Complete semantics.

11.86 MV2_USE_RoCE

This parameter enables the use of RDMA over Ethernet for MPI communication. The underlying HCA and network must support this feature.

11.87 MV2_DEFAULT_GID_INDEX

In RoCE mode, this parameter allows to choose non-default GID index in loss-less ethernet setup using VLANs

11.88 MV2_USE_RDMA_CM

This parameter enables the use of RDMA CM for establishing the connections. The library has to be built using the flag -DRDMA_CM for using this feature.

11.89 MV2_RDMA_CM_CONF_FILE_PATH

This parameter is to specify the path to mv2.conf file. If this is not given, then it searches in the default location /etc/mv2.conf

11.90 MV2_USE_RDMA_FAST_PATH

Setting this parameter enables MVAPICH2 to use adaptive RDMA fast path features for OFA-IB-CH3 interface and static RDMA fast path features for uDAPL-CH3 interface.

11.91 MV2_USE_RDMA_ONE_SIDED

Setting this parameter allows MVAPICH2 to use optimized one sided implementation based RDMA operations.

11.92 MV2_USE_RING_STARTUP

Setting this parameter enables MVAPICH2 to use ring based start up.

11.93 MV2_USE_SHARED_MEM

Use shared memory for intra-node communication.

11.94 MV2_USE_SHM_ONE_SIDED

Setting this parameter allocates the buffers requested using MPI_Alloc_mem in shared memory. One-sided communication over windows created in this memory is optimized. The use of optimization hint for MPI_Alloc_mem described in Section 6.7 is recommended over setting this parameter in order to take advantage of the optimized intra-node one-sided communication.

11.95 MV2_USE_SHMEM_ALLREDUCE

This parameter can be used to turn off shared memory based MPI_Allreduce for OFA-IB-CH3 over IBA by setting this to 0.

11.96 MV2_USE_SHMEM_BARRIER

This parameter can be used to turn off shared memory based MPI_Barrier for OFA-IB-CH3 over IBA by setting this to 0.

11.97 MV2_USE_SHMEM_BCAST

This parameter can be used to turn off shared memory based MPI_Bcast for OFA-IB-CH3 over IBA by setting this to 0.

11.98 MV2_USE_SHMEM_COLL

Use shared memory for collective communication. Set this to 0 for disabling shared memory collectives.

11.99 MV2_USE_SHMEM_REDUCE

This parameter can be used to turn off shared memory based MPI_Reduce for OFA-IB-CH3 over IBA by setting this to 0.

11.100 MV2_USE_SRQ

Setting this parameter enables MVAPICH2 to use shared receive queue.

11.101 MV2_GATHER_SWITCH_PT

We use different algorithms depending on the system size. For small system sizes (up to 386 cores), we use the “2-level” algorithm following by the “Direct” algorithm. For medium system sizes (up to 1k), we use “Binomial” algorithm following by the “Direct” algorithm. Users can set the switching point between algorithms using the run-time parameter MV2_GATHER_SWITCH_PT.

11.102 MV2_SCATTER_SMALL_MSG

When the system size is lower than 512 cores, we use the “Binomial” algorithm for small message sizes. MV2_SCATTER_SMALL_MSG allows the users to set the threshold for small messages.

11.103 MV2_SCATTER_MEDIUM_MSG

When the system size is lower than 512 cores, we use the “2-level” algorithm for medium message sizes. MV2_SCATTER_MEDIUM_MSG allows the users to set the threshold for medium messages.

11.104 MV2_USE_TWO_LEVEL_GATHER

Use the two-level multi-core-aware algorithm for the MPI_Gather operation. If this parameter is set to 0 at run-time, the two-level algorithm will not be invoked.

11.105 MV2_USE_TWO_LEVEL_SCATTER

Use the two-level multi-core-aware algorithm for the MPI_Scatter operation. If this parameter is set to 0 at run-time, the two-level algorithm will not be invoked.

11.106 MV2_USE_XRC

Use the XRC InfiniBand transport available since Mellanox ConnectX adapters. This features requires OFED version later than 1.3. It also automatically enables SRQ and ON-DEMAND connection management. Note that the MVAPICH2 library needs to have been configured with –enable-xrc=yes to use this feature.

11.107 MV2_VBUF_POOL_SIZE

The number of vbufs in the initial pool. This pool is shared among all the connections.

11.108 MV2_VBUF_SECONDARY_POOL_SIZE

The number of vbufs allocated each time when the global pool is running out in the initial pool. This is also shared among all the connections.

11.109 MV2_VBUF_TOTAL_SIZE

The size of each vbuf, the basic communication buffer of MVAPICH2. For better performance, the value of MV2_IBA_EAGER_THRESHOLD should be set the same as MV2_VBUF_TOTAL_SIZE.

11.110 MV2_SMP_EAGERSIZE

This parameter defines the switch point from Eager protocol to Rendezvous protocol for intra-node communication. Note that this variable can be set with suffixes such as ‘K’/‘k’, ‘M’/‘m’ or ‘G’/‘g’ to denote Kilobyte, Megabyte or Gigabyte respectively.

11.111 MV2_SMPI_LENGTH_QUEUE

This parameter defines the size of shared buffer between every two processes on the same node for transferring messages smaller than or equal to MV2_SMP_EAGERSIZE. Note that this variable can be set with suffixes such as ‘K’/‘k’, ‘M’/‘m’ or ‘G’/‘g’ to denote Kilobyte, Megabyte or Gigabyte respectively.

11.112 MV2_SMP_NUM_SEND_BUFFER

This parameter defines the number of internal send buffers for sending intra-node messages larger than MV2_SMP_EAGERSIZE.

11.113 MV2_SMP_SEND_BUF_SIZE

This parameter defines the packet size when sending intra-node messages larger than MV2_SMP_EAGERSIZE.

11.114 MV2_USE_HUGEPAGES

Set this to 0, to not use any HugePages.

11.115 MV2_HYBRID_ENABLE_THRESHOLD

This defines the threshold for enabling Hybrid communication using UD and RC/XRC. When the size of the job is greater than or equal to the threshold value, Hybrid mode will be enabled. Otherwise, it uses default RC/XRC connections for communication.

11.116 MV2_HYBRID_MAX_RC_CONN

Maximum number of RC or XRC connections created per process. This limits the amount of connection memory and prevents HCA QP cache thrashing.

11.117 MV2_UD_PROGRESS_TIMEOUT

Time (usec) until ACK status is checked (and ACKs are sent if needed). To avoid unnecessary retries, set this value less than MV2_UD_RETRY_TIMEOUT. It is recommended to set this to 1/10 of
MV2_UD_RETRY_TIMEOUT.

11.118 MV2_UD_RETRY_TIMEOUT

Time (usec) after which an unacknowledged message will be retried

11.119 MV2_UD_RETRY_COUNT

Number of retries of a message before the job is aborted. This is needed in case of HCA fails.

11.120 MV2_USE_UD_HYBRID

Set this to Zero, to disable UD transport in hybrid configuration mode.

11.121 MV2_USE_ONLY_UD

Set this to 1, to enable only UD transport in hybrid configuration mode. It will not use any RC/XRC connections in this mode.

11.122 MV2_USE_UD_ZCOPY

Whether or not to use the zero-copy transfer mechanism to transfer large messages on UD transport.

11.123 MV2_USE_LIMIC_GATHER

If this flag is set to 1, we will use intra-node Zero-Copy MPI_Gather designs, when the library has been configured to use LiMIC2.

11.124 MV2_USE_MCAST

Set this to 1, to enable hardware multicast support in collective communication

11.125 MV2_MCAST_NUM_NODES_THRESHOLD

This defines the threshold for enabling multicast support in collective communication. When MV2_USE_MCAST is set to 1 and the number of nodes in the job is greater than or equal to the threshold value, it uses multicast support in collective communication

11.126 MV2_USE_CUDA

set this to One. to enable support for communication with GPU device buffers.

11.127 MV2_CUDA_BLOCK_SIZE

The chunk size used in large message transfer from device memory to host memory. The other suggested values for this parameter are 131072 and 524288.

11.128 MV2_CUDA_IPC

This enables intra-node GPU-GPU communication using IPC feature available from CUDA 4.1

11.129 MV2_CUDA_SMP_IPC

This enables an optimization for short message GPU device-to-device communication using IPC feature available from CUDA 4.1

12 MVAPICH2 Parameters (OFA-IB-Nemesis Interface)

12.1 MV2_DEFAULT_MAX_SEND_WQE

This specifies the maximum number of send WQEs on each QP. Please note that for Gen2 and Gen2-iWARP, the default value of this parameter will be 16 if the number of processes is larger than 256 for better memory scalability.

12.2 MV2_DEFAULT_MAX_RECV_WQE

This specifies the maximum number of receive WQEs on each QP (maximum number of receives that can be posted on a single QP).

12.3 MV2_DEFAULT_MTU

The internal MTU size. For Gen2, this parameter should be a string instead of an integer. Valid values are: IBV_MTU_256, IBV_MTU_512, IBV_MTU_1024, IBV_MTU_2048, IBV_MTU_4096.

12.4 MV2_DEFAULT_PKEY

Select the partition to be used for the job.

12.5 MV2_IBA_EAGER_THRESHOLD

This specifies the switch point between eager and rendezvous protocol in MVAPICH2. For better performance, the value of MV2_IBA_EAGER_THRESHOLD should be set the same as MV2_VBUF_TOTAL_SIZE.

12.6 MV2_IBA_HCA

This specifies the HCA to be used for performing network operations.

12.7 MV2_INITIAL_PREPOST_DEPTH

This defines the initial number of pre-posted receive buffers for each connection. If communication happen for a particular connection, the number of buffers will be increased to
RDMA_PREPOST_DEPTH.

12.8 MV2_MAX_INLINE_SIZE

This defines the maximum inline size for data transfer. Please note that the default value of this parameter will be 0 when the number of processes is larger than 256 to improve memory usage scalability.

12.9 MV2_NDREG_ENTRIES

This defines the total number of buffers that can be stored in the registration cache. It has no effect if MV2_USE_LAZY_MEM_UNREGISTER is not set. A larger value will lead to less frequent lazy de-registration.

12.10 MV2_NUM_RDMA_BUFFER

The number of RDMA buffers used for the RDMA fast path. This fast path is used to reduce latency and overhead of small data and control messages. This value will be ineffective if MV2_USE_RDMA_FAST_PATH is not set.

12.11 MV2_NUM_SA_QUERY_RETRIES

Number of times the MPI library will attempt to query the subnet to obtain the path record information before giving up.

12.12 MV2_PREPOST_DEPTH

This defines the number of buffers pre-posted for each connection to handle send/receive operations.

12.13 MV2_RNDV_PROTOCOL

The value of this variable can be set to choose different Rendezvous protocols. RPUT (default RDMA-Write) RGET (RDMA Read based), R3 (send/recv based).

12.14 MV2_R3_THRESHOLD

The value of this variable controls what message sizes go over the R3 rendezvous protocol. Messages above this message size use MV2_RNDV_PROTOCOL.

12.15 MV2_R3_NOCACHE_THRESHOLD

The value of this variable controls what message sizes go over the R3 rendezvous protocol when the registration cache is disabled (MV2_USE_LAZY_MEM_UNREGISTER=0). Messages above this message size use MV2_RNDV_PROTOCOL.

12.16 MV2_SRQ_LIMIT

This is the low water-mark limit for the Shared Receive Queue. If the number of available work entries on the SRQ drops below this limit, the flow control will be activated.

12.17 MV2_SRQ_SIZE

This is the maximum number of work requests allowed on the Shared Receive Queue.

12.18 MV2_STRIPING_THRESHOLD

This parameter specifies the message size above which we begin the stripe the message across multiple rails (if present).

12.19 MV2_USE_BLOCKING

Setting this parameter enables mvapich2 to use blocking mode progress. MPI applications do not take up any CPU when they are waiting for incoming messages.

12.20 MV2_USE_LAZY_MEM_UNREGISTER

Setting this parameter enables mvapich2 to use memory registration cache.

12.21 MV2_USE_RDMA_FAST_PATH

Setting this parameter enables mvapich2 to use adaptive rdma fast path features for Gen2 interface and static rdma fast path features for uDAPL interface.

12.22 MV2_USE_SRQ

Setting this parameter enables mvapich2 to use shared receive queue.

12.23 MV2_VBUF_POOL_SIZE

The number of vbufs in the initial pool. This pool is shared among all the connections.

12.24 MV2_VBUF_SECONDARY_POOL_SIZE

The number of vbufs allocated each time when the global pool is running out in the initial pool. This is also shared among all the connections.

12.25 MV2_VBUF_TOTAL_SIZE

The size of each vbuf, the basic communication buffer of MVAPICH2. For better performance, the value of MV2_IBA_EAGER_THRESHOLD should be set the same as MV2_VBUF_TOTAL_SIZE.

12.26 MV2_RUN_THROUGH_STABILIZATION

This enables run through stabilization support to handle the process failures. This is valid only with Hydra process manager with –disable-auto-cleanup flag.

13 MPIRUN_RSH Parameters

13.1 MV2_COMM_WORLD_LOCAL_RANK

The local rank of a process on a node within its job. The local rank ranges from 0,1 ... N-1 on a node with N processes running on it.

13.2 MV2_COMM_WORLD_LOCAL_SIZE

The number of ranks from this job that are running on this node.

13.3 MV2_COMM_WORLD_RANK

The MPI rank of this process in current MPI job

13.4 MV2_COMM_WORLD_SIZE

The number of processes in this MPI job’s MPI_Comm_World.

13.5 MV2_FASTSSH_THRESHOLD

Number of nodes beyond which to use hierarchical ssh during start up. This parameter is only relevant for mpirun_rsh based start up. Note that unlike most other parameters described in this section, this is an environment variable that has to be set in the run time environment (for e.g. through export in the bash shell).

13.6 MV2_NPROCS_THRESHOLD

Number of nodes beyond which to use file-based communication scheme in the hierarchical ssh during start up. This parameter is only relevant for mpirun_rsh based start up.

13.7 MV2_MPIRUN_TIMEOUT

The number of seconds after which mpirun_rsh aborts job launch. Note that unlike most other parameters described in this section, this is an environment variable that has to be set in the run time environment (for e.g. through export in the bash shell).

13.8 MV2_MT_DEGREE

The degree of the hierarchical tree used by mpirun_rsh. By default mpirun_rsh uses a value that tries to keep the depth of the tree to 4. Note that unlike most other parameters described in this section, this is an environment variable that has to be set in the run time environment (for e.g. through export in the bash shell).

13.9 MPIEXEC_TIMEOUT

Set this to limit, in seconds, of the execution time of the mpi application. This overwrites the
MV2_MPIRUN_TIMEOUT parameter.

13.10 MV2_DEBUG_FORK_VERBOSE

Set the verbosity level of the debug output for the process management operations (fork, waitpid, kill, ...) of mpirun_rsh and mpispawn processes. The value 0 disables any debug output, a value of 1 enables a basic debug output, and a value of 2 enables a more verbose debug output.

Note: All debug output is disabled when MVAPICH2 is configured with the --enable-fast=ndebug option.