Application Performance of MVAPICH2-GDR
|CPU Model||CPU Core Info||Memory||IB Card||OS||OFED||GPU||CUDA|
|Intel E5-2630||2x8 @ 2.6Ghz||64GB||Mellanox Connect-IB||Redhat 6.4||MOFED 2.1||NVIDIA Tesla K20||CUDA 6.0|
The performance numbers were taken with the following flags:
- gdrcopy enabled: MV2_USE_CUDA=1 MV2_IBA_HCA=mlx5_0 MV2_IBA_EAGER_THRESHOLD=32768 MV2_VBUF_TOTAL_SIZE=32768 MV2_USE_GPUDIRECT_LOOPBACK_LIMIT=32768 MV2_USE_GPUDIRECT_GDRCOPY=1 MV2_USE_GPUDIRECT_GDRCOPY_LIMIT=16384
These numbers were taken on Wilkes system. We would like to thank the University of Cambridge and the staff members of Wilkes for providing us access to this system.