[mvapich-discuss] Question with bandwidth tests of OSU Microbenchmarks
Junchao Zhang
junchao.zhang at gmail.com
Wed Dec 13 10:53:09 EST 2017
Hello,
I used the latest osu_mbw_mr and measured bandwidth between two sockets
of a NUMA node. I found the bandwidth for middle-sized messages was very
large, even bigger than memory bandwidth of the node. I don't understand
why. Shouldn't it be bound by QPI bandwidth of the node?
Also, I find in osu_mbw_mr.c a receiver issues multiple MPI_Irecv's, but
they all have the same r_buf, which means the received data is overlapped.
Is it a bug or done intentionally?
Here is my test result on an Intel Xeon Haswell node with 2 sockets, and
32 cores. The QPI bandwidth is 38.4GB/s. The STREAM copy bandwidth is
110GB/s. I used mvapich2-2.1-intel-13.1. When running the code, I put first
16 ranks on socket 0 and the next 16 on socket 1. Could you shed some light
on that? Thank you.
# OSU MPI Multiple Bandwidth / Message Rate Test v5.4.0
# [ pairs: 16 ] [ window size: 64 ]
# Size MB/s Messages/s
1 67.89 67888521.24
2 140.70 70348753.88
4 273.76 68440922.18
8 558.12 69764549.69
16 984.26 61516508.73
32 1959.77 61242938.77
64 2901.28 45332460.41
128 5830.29 45549103.35
256 12216.54 47720864.66
512 25322.88 49458754.12
1024 40257.18 39313651.61
2048 68684.93 33537564.96
4096 79012.29 19290110.86
8192 88479.92 10800771.52
16384 112174.38 6846581.09
32768 125500.76 3829979.11
65536 134492.98 2052200.00
131072 136100.17 1038361.91
262144 132753.53 506414.51
524288 125918.77 240171.00
1048576 74015.27 70586.46
2097152 30469.80 14529.13
4194304 27742.19 6614.25
--Junchao Zhang
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20171213/d0a17040/attachment-0001.html>
More information about the mvapich-discuss
mailing list