I finally have opportunity to run the imb-3.2 benchmark over myrinet I am running in a cluster of 16 node Xservers connected with myrinet 15 of them are 8core ones and the last one is a 4 cores one. Having a limit of 124 process
I have run the test with the bynode option so from the 2 to the 16 process test is always running 1 process by node.
the following test pingpong, pingping, sendrecv, exchange presents a strong drop in performance with the 64k packet size.
any idea where I should look for the cause.
Ricardo
On Fri, Mar 20, 2009 at 7:32 PM, Ricardo Fernández-Perea
<rfernandezperea@gmail.com> wrote:
It is the F-2M but I think for inter-node communication should be equivalents.
I have not run and MPI pingpong benchmark yet.
The truth is I have a 10 days travel coming next week and I thought I can take some optimization "light reading" with me.
so I know what I must look for when I came back.
Ricardo
On Fri, Mar 20, 2009 at 5:10 PM, Scott Atchley
<atchley@myri.com> wrote:
On Mar 20, 2009, at 11:33 AM, Ricardo Fernández-Perea wrote:
This are the results initially
Running 1000 iterations.
Length Latency(us) Bandwidth(MB/s)
0 2.738 0.000
1 2.718 0.368
2 2.707 0.739
<snip>
1048576 4392.217 238.735
2097152 8705.028 240.913
4194304 17359.166 241.619
with export MX_RCACHE=1
Running 1000 iterations.
Length Latency(us) Bandwidth(MB/s)
0 2.731 0.000
1 2.705 0.370
2 2.719 0.736
<snip>
1048576 4265.846 245.807
2097152 8491.122 246.982
4194304 16953.997 247.393
Ricardo,
I am assuming that these are PCI-X NICs. Given the latency and bandwidth, are these "D" model NICs (see the top of the mx_info output)? If so, that looks about as good as you can expect.
Have you run Intel MPI Benchmark (IMB) or another MPI pingpong type benchmark?
Scott