JFS bug

[tridge at samba.org]

Lars,

 > I have heard that tridge has discovered a possible bug in JFS. I am
 > new to samba development but I have hacked on JFS since year 2000
 > (http://kerneltraffic.org/debian-hurd/dh20000607_51.html#6 among
 > others). So if there is a bug then please step forward with a bug
 > report (or a link to a report) and I will try to fix it.

I certainly suspect it is a jfs bug. See
http://samba.org/~tridge/xattr_results/ for the script to run the
benchmark, the script to produce the graphs, and the graphs of the
results showing that jfs performs far worse than could be possibly
expected.  I suspect serious lock contention.

It would be interesting to produce a graph showing how much actual
block IO is happening for jfs and other filesystems in this test. What
I suspect you will find is that jfs isn't doing much more actual block
IO per operation, but that instead it is blocking a lot waiting for
some lock. You could use vmstat to check this.

See http://samba.org/ftp/tridge/dbench/ for a copy of the benchmark
code.

I should also note that I have previously seen very good results from
jfs with earlier versions of the Linux kernel on quite similar
benchmarks. That is another reason I strongly suspect a bug rather
than a design flaw.

Cheers, Tridge