read benchmarks: ufs/zfs/ext3 raidz/raid5

[Ivan Voras]

Dan Naumov wrote:
> Another FreeBSD person on a forum I frequent did some read benchmarks
> on his system: Athlon64 3500+ with 2GB DDR2 SDRAM, a WD 250GB system
> drive, and 5 Seagate Barracuda 750GB SATA-II data drives. ZFS and UFS
> testing was done using FreeBSD 7.2-RELEASE amd64, and ext3 testing was
> done using Ubuntu Server 8.04-LTS amd64. The used disks do not support
> NCQ, so there is no "NCQ advantage" on the Linux side.
> 
> Random Access reads, 5MB chunks:
> http://virtual.tehinterweb.net/livejournal/2009-06-22_zfs_diskperf/zfs-diskperf-untuned-5mb.png
> Random Access reads, 1MB chunks:
> http://virtual.tehinterweb.net/livejournal/2009-06-22_zfs_diskperf/zfs-diskperf-untuned-1mb.png
> Random Access reads, 5MB chunks (big list):
> http://virtual.tehinterweb.net/livejournal/raid_performance/raid-diskperf-5mb-all.png
> 
> Here is the original forum discussion thread:
> http://episteme.arstechnica.com/eve/forums/a/tpc/f/96509133/m/857002910041

This looks about consistent with what I see when comparing ext3 to ZFS
and/or UFS on 7.x systems.

I have theories and hunches why IO in FreeBSD is slower than on Linux
(and for different RAID tools / GEOM classes the reasons are slightly
different) but nothing I can back with proof and measurements. I can
only confirm that it is consistently slower.

I haven't yet tried testing 8 so if you're looking for ideas for
testing, try it (disable debugging before you use 8-CURRENT). I think
ZFS has some concurrency-enhancing additions there (which will help in
case the benchmark was done with a tool testing concurrency; I don't see
what was used in the above pages).



-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 258 bytes
Desc: OpenPGP digital signature
Url : http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20090624/976dbe72/signature.pgp