Bandwidth on NASA pleiades supercomputer

Kihiro Bando

unread,

Feb 19, 2020, 9:25:55 PM2/19/20

to gasnet...@lbl.gov

Dear all,

I am testing GASNet on pleiades for which you can see the specs here. I ran the suggested tests and found out that latency is comparable but the bandwidth is half of that reported on the website.

Please find attached the supporting figures.

Do these numbers look reasonable to you?

Thank you very much.

Best,
Kihiro Bando

bandwidth_test.png

latency_test.png

Paul Hargrove

unread,

Feb 20, 2020, 12:20:27 AM2/20/20

to Kihiro Bando, gasnet-users

Kihiro,

I don't find the word "bandwidth" anywhere on the specs page you provided.

Can you please help me find the value you are comparing to when you say "half"?

You may be able to determine on your own if the bandwidth number is "reasonable" by running a MPI-based bandwidth test. We distribute one you can try. If you `make tests-mpi` in the ibv-conduit directory, one of the tests generated is `testosumpi_bw` which is a version of an OSU benchmark modified to report numbers suitable for comparison to those from GASNet's `testlarge`.

-Paul

--
You received this message because you are subscribed to the Google Groups "gasnet-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gasnet-users...@lbl.gov.
To view this discussion on the web visit https://groups.google.com/a/lbl.gov/d/msgid/gasnet-users/ed0861ce-4b46-2c98-6fdf-cab4703fabb1%40stanford.edu.

--

Paul H. Hargrove <PHHar...@lbl.gov>

Computer Languages & Systems Software (CLaSS) Group

Computer Science Department

Lawrence Berkeley National Laboratory

Kihiro Bando

unread,

Feb 20, 2020, 12:24:07 AM2/20/20

to Paul Hargrove, gasnet-users

Hi Paul,

Thank you for your response. I was referring to the tests on GASNet webpage. I guess it is not as relevant as I thought to directly compare both results.

I will try the mpi and will keep you in touch.

Thank you very much.
Kihiro

Kihiro Bando

unread,

Feb 21, 2020, 12:50:03 PM2/21/20

to Paul Hargrove, gasnet-users

Hi Paul,

I was not able to find the source file for the test. The target `tests-mpi` is only in the tests folder and when I execute `make tests-mpi` in tests, it complains about not finding the test source files (testmpiperf.c to begin with).

Kihiro

Paul Hargrove

unread,

Feb 21, 2020, 1:00:38 PM2/21/20

to Kihiro Bando, gasnet-users

Kihiro,

My apologies. I had forgotten that the test in question is in our git repository but not distributed in our tarball.

The source file can be found here .

I can be compiled and run as any other MPI program would be.

-Paul

Kihiro Bando

unread,

Feb 21, 2020, 1:08:29 PM2/21/20

to Paul Hargrove, gasnet-users

Thanks. I should run it with `mpiexec -np 2 -ppn 1 ./testosumpi_bw` on 2 nodes to have relevant comparison right? The values are then comparable with my testlarge.

Paul Hargrove

unread,

Feb 21, 2020, 1:21:55 PM2/21/20

to Kihiro Bando, gasnet-users

Yes, that's right.

This test has been modified relative to the original to report in units of MB = 2^20 (instead of 10^6).

If you divide the KB/s results from testlarge by 1024, then the two tests will be reporting comparable results.

-Paul

Kihiro Bando

unread,

Feb 21, 2020, 1:26:07 PM2/21/20

to Paul Hargrove, gasnet-users

My testlarge gave results in MB/s so I did not have to rescale the data. Thank you for your help.
Kihiro

Reply all

Reply to author

Forward