[mvapich-discuss] mvapich2-1.8ap cuda osu bandwidth test

[sreeram potluri]

Hi,

Can you give us more details about your runs:

1) the system you are running these benchmarks on, primarily the processor
architecture, the GPU(s) and the network card

2) the list of run-time parameters you are using

Sreeram Potluri

On Tue, Jan 24, 2012 at 10:21 AM, Yaoming Mu <yaomingmu at gmail.com> wrote:

> I installed mvapich2-1.8ap with cuda support, but bandwidth test
> results were very bad , compared to results published by mvapich2 team
>
> # OSU MPI-CUDA Bandwidth Test
> # Send Buffer on HOST (H) and Receive Buffer on DEVICE (D)
> # Size        Bandwidth (MB/s)
> 1                         0.12
> 2                         0.25
> 4                         0.57
> 8                         1.09
> 16                        2.19
> 32                        4.42
> 64                        8.96
> 128                      17.78
> 256                      35.42
> 512                      69.32
> 1024                    132.06
> 2048                    243.37
> 4096                    408.74
> 8192                    557.47
> 16384                   479.83
> 32768                   542.55
> 65536                   576.15
> 131072                  598.19
> 262144                  606.11
> 524288                  611.05
> 1048576                 611.86
> 2097152                 610.60
> 4194304                 607.65
>
>
>  even Host to Host is so bad if I export MV2_USE_CUDA=1
>
> # OSU MPI-CUDA Bandwidth Test
> # Send Buffer on HOST (H) and Receive Buffer on HOST (H)
> # Size        Bandwidth (MB/s)
> 1                         1.04
> 2                         2.67
> 4                         5.34
> 8                        10.52
> 16                       21.01
> 32                       41.53
> 64                       81.90
> 128                     140.18
> 256                     159.54
> 512                     271.31
> 1024                    389.32
> 2048                    485.95
> 4096                    557.83
> 8192                    585.33
> 16384                   493.49
> 32768                   553.46
> 65536                   589.91
> 131072                  608.17
> 262144                  619.25
> 524288                  622.58
> 1048576                 624.03
> 2097152                 624.59
> 4194304                 623.50
>
> But if disable cuda by export MV2_USE_CUDA=0, I got normal results
> # OSU MPI-CUDA Bandwidth Test
> # Send Buffer on HOST (H) and Receive Buffer on HOST (H)
> # Size        Bandwidth (MB/s)
> 1                         4.12
> 2                         8.42
> 4                        16.22
> 8                        33.31
> 16                       70.34
> 32                      122.98
> 64                      262.41
> 128                     503.14
> 256                     957.10
> 512                    1657.49
> 1024                   2806.88
> 2048                   5126.41
> 4096                   7364.44
> 8192                  10301.80
> 16384                 10391.13
> 32768                  9811.05
> 65536                 11492.53
> 131072                11793.38
> 262144                10714.53
> 524288                10282.15
> 1048576               10413.02
> 2097152               10124.43
> 4194304                5963.46
>
> My cuda version is 4.1rc2
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20120124/eb18da96/attachment.html