http://groups.google.com/group/zfs-fuse/msg/5fac5eaf2c7fccb8 shows some (admittedly very crude) tests I did with OpenSolaris 0906, with some very surprising performance results. In particular, read speed on an 8-disk pool seemed to drop by 50% when I set up the pool to use RAIDZ2. Can anyone shed light on why that might be the case? Thanks, -- Dave Abrahams BoostPro Computing http://www.boostpro.com
David Abrahams wrote:> http://groups.google.com/group/zfs-fuse/msg/5fac5eaf2c7fccb8 shows some > (admittedly very crude) tests I did with OpenSolaris 0906, with some > very surprising performance results. In particular, read speed on an > 8-disk pool seemed to drop by 50% when I set up the pool to use RAIDZ2. > > Can anyone shed light on why that might be the case? >You are likely disk I/O bound. Look at the disk traffic to locate your bottleneck. FWIW, if you enable compression, your benchmark will be much faster :-) -- richard
on Fri May 22 2009, Richard Elling <richard.elling-AT-gmail.com> wrote:> David Abrahams wrote: >> http://groups.google.com/group/zfs-fuse/msg/5fac5eaf2c7fccb8 shows some >> (admittedly very crude) tests I did with OpenSolaris 0906, with some >> very surprising performance results. In particular, read speed on an >> 8-disk pool seemed to drop by 50% when I set up the pool to use RAIDZ2. >> >> Can anyone shed light on why that might be the case? >> > > You are likely disk I/O bound.Even if I were, that shouldn''t account for such a dramatic difference AFAICT. Shouldn''t this hold: raidz2 / flatzfs = (raidpoolsize - raidparitybits) / (flatpoolsize) ?? In my case raidpoolsize=flatpoolsize=8 and paritybits=2, so the ratio should be 0.75, not 0.5. Or am I missing something?> Look at the disk traffic to locate your bottleneck.How does one look at the disk traffic?> FWIW, if you enable compression, your benchmark will be > much faster :-)Of course, but writing a string of zeroes to a compressed filesystem only tests the compression speed, which is of little or no interest to me. Incidentally, I have some speed tests of a compressed zfs here: http://groups.google.com/group/zfs-fuse/browse_thread/thread/b55a25ea432b416 but as the message implies, it''s hard to know whether those results mean anything. -- Dave Abrahams BoostPro Computing http://www.boostpro.com
> How does one look at the disk traffic?iostat -xce 1 > OpenSolaris, raidz2 across 8 7200 RPM SATA disks: > 17179869184 bytes (17 GB) copied, 127.308 s, 135 MB/s > OpenSolaris, "flat" pool across the same 8 disks: > 17179869184 bytes (17 GB) copied, 61.328 s, 280 MB/s one raidz2 set of 8 disks can''t be faster than the slowest disk in the set as its one vdev... I would have expected the 8 vdev set to be 8x faster than the single raidz[12] set, but like Richard said, there is another bottle neck in there that iostat will show. Rob