I''m seeing a performance regression for Writes in 1.6.5 relative to 1.6.4.3. However, I''m also seeing a performance improvement for reads. I''m running over Mellanox IB fabric. Is anyone else seeing this? Here are some numbers for my HW: 1.6.4.3 + ofed 1.3: Writes 390 MB/s, Reads 369 MB/s 1.6.5 + ofed 1.3.1: Writes 366 MB/s, Reads 437 MB/s Has anyone else taken performance measurements? Roger Spellman Staff Engineer Terascala, Inc. 508-588-1501 www.terascala.com <http://www.terascala.com/> -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.lustre.org/pipermail/lustre-discuss/attachments/20080711/3b2ac11c/attachment.html
On Jul 11, 2008 16:46 -0400, Roger Spellman wrote:> I''m seeing a performance regression for Writes in 1.6.5 relative to > 1.6.4.3. However, I''m also seeing a performance improvement for reads. > I''m running over Mellanox IB fabric. Is anyone else seeing this? > > > > Here are some numbers for my HW: > 1.6.4.3 + ofed 1.3: Writes 390 MB/s, Reads 369 MB/s > > 1.6.5 + ofed 1.3.1: Writes 366 MB/s, Reads 437 MB/s > > Has anyone else taken performance measurements?There was one similar thread on the list... Did you try turning off the checksumming? That is again on by default in 1.6.5, with an improved algorithm, but can be disabled at runtime on the clients with: lctl set_param osc.*.checksums=0 or permanently on the MGS (MDS) with: lctl conf_param {fsname}.osc.checksums=0 Cheers, Andreas -- Andreas Dilger Sr. Staff Engineer, Lustre Group Sun Microsystems of Canada, Inc.
> Did you try turning off the > checksumming? That is again on by default in 1.6.5, with an improved > algorithm ...Thanks, Andreas. I see that this is the cause of my problem. This did not make it into the release notes, as far as I can see.