Re: 2.6.12-mm1 & 2K lun testing (JFS problem ?)

[William Lee Irwin III <wli@holomorphy.com>]

On Tue, Jun 21, 2005 at 05:34:54PM -0700, Badari Pulavarty wrote:
> Hi Andrew & Shaggy,
> Here is the summary of 2K lun testing on 2.6.12-mm1.
> When I tune dirty ratios and CFQ queue depths, things
> seems to be running fine.
> 	echo 20 > /proc/sys/vm/dirty_ratio
> 	echo 20 > /proc/sys/vm/overcommit_ratio
> 	echo 4 > /sys/block/<device>/queue/nr_requests
> But, I am running into JFS problem. I can't kill my
> "dd" process. They all get stuck in:
> (I am going to try ext3).

If you could get unabridged profiling data for raw vs. fs (so it can
be properly sorted) I would be interested in that. Early indications
were large amounts of time spent in shrink_zone(), obtained by
re-sorting the truncated profile listings. It indicated the time spent
in shrink_zone() was 26.3 times as much time spent in default_idle().
Typically copying to and from userspace are enormous overheads, but
aren't observable in the truncated/mis-sorted profiles, which calls
them into question, barring unreported usage of O_DIRECT. There are
also no totals reported, which are helpful for interpreting realtime
behavior.


-- wli