On Mon, May 18, 2009 at 10:52 AM, Sufficool, Stanley <[email protected]> wrote: > IIRC, The SRP Target code has many context switches that throttle > performance at higher thread counts.
Can anything be done to reduce the context switches? Is there, for example, one thread on the target per user thread that may best be pinned? Thanks, Chris > >> -----Original Message----- >> From: Chris Worley [mailto:[email protected]] >> Sent: Monday, May 18, 2009 9:21 AM >> To: OpenIB; scst-devel >> Subject: [Scst-devel] SRP aggregate bandwidth decreasing as >> threads increase >> >> >> I'm seeing peak performance at ~4 threads (1.6GB/s), w/ >> threads >10 I'm seeing aggregate performance drop >> significantly. This is not a drive issue: locally, the >> drives get best performance >~32 threads, and maintain their >> aggregate way beyond that. >> >> Is there any tunable parameter or source code change in the >> initiator or target code that would effect performance with a >> high thread count? >> >> Thanks, >> >> Chris _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
