Re: ZFS performance weirdness on a a Netra X1

Tim Bradshaw <tfb@xxxxxxxx> wrote:
I have a Netra X1 with a pair of 120GB (non-Sun) disks, and 1152MB of
memory (so enough, I hope). It's running 10 6/06, and the disks are
split between SVM and: /, /var and swap are UFS filesystems on SVM
mirrors, and the rest is taken up by zfs:

That way ZFS cannot enable the write cache on the disks. But these
are IDE disks and on non-Sun IDE disks the write cache is normally enabled
by default. So this type of setup should cause the performance
degradation.

with resulting write rates of about 3MB/s through zfs, and about 14
through UFS and SVM, onto the same pair of disks, obviously. 3MB/s is
pretty crippling. Obviously these aren't realistic FS benchmarks, but
they are indicative of something.

How did you set up the individual zfs filesystems? Did you change the
checksum algorithm to sha256? This will kill performance - especially on
such a slow machine. What is the output of "zfs get all export/zones/ts"
(or the other names)?

I have a similar setup (two 120 GB disks mirrored) and did also a small
write test (1 GB in total):

timex dd if=/dev/zero of=bigfile bs=1024k count=1024
1024+0 Datensätze in
1024+0 Datensätze aus

real 42.15
user 0.00
sys 2.26

So I achieved a write performance of ~25 MB/s. Although my machine is much
faster than a Netra X1, this shouldn't matter much: During the test the
system was ~95% idle.

--
Daniel

.

Relevant Pages