zfs very poor performance compared to ufs due to lack of cache?

Andriy Gapon avg at freebsd.org
Tue Sep 21 14:23:14 UTC 2010

on 21/09/2010 16:53 Steven Hartland said the following:
> That's what I thought you where saying. Is there a test you would suggest to confirm
> either way more accurately?

Perhaps you can try the test scenario that you described and monitor parameters
suggested by Wiktor in this thread.

That is, have two large files and set arc max size such that one of them can fit
in ARC readily, but two of them won't fit by a large margin.  Make sure that
remaining RAM is large enough to hold both files in page cache.

1. sendfile one file, then the other
2. record kstat.zfs.misc.arcstats values
3. sendfile the first file again
4. record kstat.zfs.misc.arcstats values

If the first file data was re-used from page cache, then you won't see much
changes in kstat.zfs.misc.arcstats.  If it had to be taken from ARC or from disk,
then either ARC hits or ARC misses will grow noticeably.

Make sure to not have any parallel activity that could affect kstat.zfs.misc.arcstats.

I think kstat.zfs.misc.arcstats.hits and kstat.zfs.misc.arcstats.misses are two
primary indicators in this test.

Andriy Gapon

More information about the freebsd-fs mailing list