Ohio State University

Hybrid MPI+OpenSHMEM 2D-Heat Benchmark | Performance | Network-Based Computing Laboratory

Hybrid MPI+OpenSHMEM 2D-Heat Benchmark (05/06/13)

  • Experimental Testbed: Each compute node has eight Intel Xeon cores running at 2.67 GHz with 12 MB L3 cache. The cores are organized as two sockets with four cores per socket. Each node has 12GB of memory and is equipped with MT26428 QDR ConnectX HCAs (36 Gbps data rate) with PCI-Ex Gen2 interfaces. The nodes are interconnected using 171- port Mellanox QDR switch. The operating system used is Red Hat Enterprise Linux Server release 5.4 (Santiago), with kernel version 2.6.32-71.el6 and OpenFabrics version 1.5.3.
  • 2D-Heat Benchmark is an OpenSHMEM benchmark used for modeling heat condution. It uses Gauss-Seidel (GS) method for the modeling. The GS kernel is repeated until the standard deviation between adjacent 2D matrices is less than a predefined convergence value. The benchmark uses both point-to-point and collective communication calls. The hybrid design (MPI+OpenSHMEM) uses MPI collectives instead of OpenSHMEM collectives. The hybrid design offers significant performance improvement as compared to the pure OpenSHMEM implementation.
  • The bar graph depicts the performance improvements that can be obtained for the 2D-Heat Benchmark by using hybrid MPI+OpenSHMEM design usign MVAPICH2-X, as compared to pure OpenSHMEM implementation. The input size for the benchmark is an 8K x 8K matrix.