[mvapich-discuss] Disabling Cache for OSU Micro-Benchmarks

[Subramoni, Hari]

Hi, Mohammad.

OSU microbenchmarks does not have a method to flush the L2/L3 caches between iterations at this point in time.

Are you performing intra-node (two processes in one physical node) or inter-node (one process per node, between two physical nodes) tests? If it is an inter-node test, it is unlikely that caching can lead to a large difference in performance.

OSU microbenchmarks reports numbers in million bytes per second. Not megabytes per second. This is another thing can could explain the difference.

Thx,
Hari.

From: mvapich-discuss-bounces at cse.ohio-state.edu On Behalf Of Mohamad Othman Sindi
Sent: Monday, November 12, 2018 12:25 PM
To: mvapich-discuss at cse.ohio-state.edu <mvapich-discuss at mailman.cse.ohio-state.edu>
Subject: [mvapich-discuss] Disabling Cache for OSU Micro-Benchmarks

Dear All,

I just started using the OSU micro-benchmarks to do some network benchmarks using osu_bw and osu_latency.

The benchmarks are compiled with MPICH (not MVAPICH) as I am running on Ethernet.

When using osu_bw I am getting bandwidth results that are larger than the theoretical 1G and 10G networks that I am running on.

I am guessing this might be due to caching.

Is there a way to disable caching when running osu_bw?

Previously I've used the Intel MPI Benchmarks (IMB) to do similar tests and they had an option "-off_cache" to disable caching when launching the benchmark and I am just wondering if there's something similar with the OSU benchmarks.

Many thanks for your help.

Sincerely,

Mohamad




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20181112/fa086e10/attachment.html>