On Sun, Apr 12, 2009 at 11:23:37AM -0400, Jim Mauro wrote: > Thank you Roland. I will try and get an nv110 build in-house > and reproduce this. Your dd test after reboot is a single threaded > sequential read, so I still don't get how disabling prefetch yields > a 15X bandwidth increase. > > I appreciate the update. > > Thanks, > /jim > > > roland wrote: >> Hello Jim, >> >> i double checked again - but it`s like i told: >> >> echo zfs_prefetch_disable/W0t1 | mdb -kw >> >> fixes my problem.
The last time I saw a problem like this, it was caused by a flaky SATA driver. Disabling prefetch hid the problem, as did reducing the number of simultaneous commands ZFS could issue to the device. This is the most complete thread on this topic that I've been able to find: http://mail.opensolaris.org/pipermail/zfs-discuss/2007-May/011083.html Although it goes on at length, it describes a technique that you can use to determine whether your hardware is having trouble with simultaneous outstanding I/Os. If you turn prefetch back on, and set zfs_vdev_{min,max}_pending=1 and verify that you only have a single outstanding I/O (use iostat for this), and performance is good, then the problem is likely with your hardware. http://mail.opensolaris.org/pipermail/zfs-discuss/2007-May/011168.html -j _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
