ZFS Performance
Gordan Bobic
an unusual performance pattern. The setup is a zpool of 5 SATA disks in
raidz2, with dedup=on. The data test set contains about 50GB of files
mostly between 1MB and 10MB in size, with very little duplication in the
data set (low single figure percentage figures).
To begin with, the performance is passably good at about 60MB/s (not
great for 3+2 SATA disks, though). By the time there is about 5GB of
data copied to ZFS, performance starts to slow down quite rapidly up to
the point where there is about 8GB copied. At this point, the write
speed goes down to about 10MB/s. It then seems to stay there and doesn't
deteriorate further. The CPU usage at this point is about 20-25%.
iostat shows that the tps figures on the disk are fairly static
throughout the operation, but the number of reads on the disks in the
zpool goes up.
My guess is that these reads are caused by checksum lookups for
deduplication. Is that the case?
Are there any tuning parameters that can be applied to optimize this?
Considering the machine has 8GB of RAM in it, would expect the hashes to
be cached and not require read lookups at all.
Gordan
Gordan Bobic
Forgot to mention - doing:
# zfs set dedup=off myzpool
gets the performance back to 60MB/s. The CPU usage goes up to 35-40%.
Interestingly, the iowait time goes up from 0% to about 25% at the same
time.
What adjustments can be made to the default configuration to improve on
this?
Gordan