Announcing: ZFS-fuse 0.7.0
#37 — Help:Is zfs-fuse's performance is not good?
| State | Resolved |
|---|---|
| Version: |
—
|
| Area | Functionality |
| Issue type | Feature |
| Severity | Medium |
| Submitted by | superpopb2b |
| Submitted on | Apr 25, 2010 |
| Responsible | Seth Heeren |
| Target release: |
—
|
Last modified on
May 06, 2010
by
Seth Heeren
Hi all, I have tested zfs-fuse by the benchmark iozone. But the result is not as good. I created a raidz zpool by using 6 SAS disk. The sequential writting performance of zfs is 36MB/S or so. After that I used the same disks and created raid5 array disk, and formatted to ext3 file system, whose performance is 120MB/S. But we see in the Sun's documents, the ZFS's performance should be better than other file system such ext2 or ext3. I would like to know why my testing result was on the contrary of Sun's documents. I guess it is because the fuse version in Linux is much slower than the non-fuse version in Solaris, but I'm not sure. Is someone know why and help me? Thank you .
PS: my ZFS-fuse version is 0.5.0.
PS: my ZFS-fuse version is 0.5.0.
Added by
Seth Heeren
on
Apr 25, 2010 09:17 AM
Issue state:
unconfirmed → open
Responsible manager:
(UNASSIGNED) → sgheeren
Yes, poor write performance is a known issue with zfs-fuse.
Mandatory: upgrade to zfs-fuse 0.6.0-critical (various packages exist in various places on the net. Be sure to check for the version).
This version comes with heaps of performance improvements. Be sure to edit /etc/zfs/zfsrc (sample in contrib/ of the git repo).
Do _not_ upgrade the pool (or fs) versions if you wish to retain the possibility to go back to 0.5.0
The bad write performance manifests itself with a redundant vdev other than mirror. Read performance matches OpenSolaris ZFS (sometimes surpassing it by a tiny margin).
Running a zil/cache on ssd can help alleviate the pattern.
Also, the usual optimizations apply. You can safely disable the elevator scheduler (linux kernel option or echo noop > /sys/block/sdx/queue/scheduler)
Mounting noatime (zfs set -r noatime=off mypoolname)
Mandatory: upgrade to zfs-fuse 0.6.0-critical (various packages exist in various places on the net. Be sure to check for the version).
This version comes with heaps of performance improvements. Be sure to edit /etc/zfs/zfsrc (sample in contrib/ of the git repo).
Do _not_ upgrade the pool (or fs) versions if you wish to retain the possibility to go back to 0.5.0
The bad write performance manifests itself with a redundant vdev other than mirror. Read performance matches OpenSolaris ZFS (sometimes surpassing it by a tiny margin).
Running a zil/cache on ssd can help alleviate the pattern.
Also, the usual optimizations apply. You can safely disable the elevator scheduler (linux kernel option or echo noop > /sys/block/sdx/queue/scheduler)
Mounting noatime (zfs set -r noatime=off mypoolname)
Added by
Seth Heeren
on
Apr 25, 2010 09:26 AM
oh and check whether your distribution makes it easy to upgrade to fuse 2.8.x if necessary. Fuse 2.8.x enables big_writes which could (depending on the load type) increase throughput dramatically
Added by
superpopb2b
on
Apr 25, 2010 10:34 AM
Thanks very much! I will try to upgrade the zfs-fuse and the fuse version and test again.
Added by
Seth Heeren
on
May 06, 2010 05:55 AM
Issue state:
open → resolved
closing
