elevator=as related to my 2.6 performance issues

[Antti Lankila]

I whined a couple of weeks back with stalling USB and PS/2 mice... I seem to
have accidentally found a workaround around the issue.

I switched to elevator=deadline the other day and found that it solved my
issues with stalling USB mice that I'm seeing on 3 computers. Overall the
system behaves very responsively now during IDE load, none of the bad
behaviour such as:

- system is something like 80 % idle but still no events from USB or PS/2
  mice appear in /dev/psaux, thus mouse doesn't move for prolonged periods.
  I've measured the mouse to stop moving for up to 10 seconds.

- when apt-get dist-upgrade or updatedb is running, games like ut2004 start
  to stall so much everything gets unplayable. You don't expect
  /usr/bin/find to cause too much load. On the contrary, you would expect it
  to be hardly noticeable.

I've seen this on 3 computers now, which have been based on KT266, KT333 and
Nforce2. Kernels have been 2.6.3-1-k7 (Debian sarge) and 2.6.5 (custom
compiled).

I tried elevator=deadline because I got tired of the unacceptable
performance I was seeing from 2.6 and tried to come up with some kind of
numbers with bonnie++ to prove that 'hey, look there, that's the problem'.
Bonnie++ numbers with 2.6 AS looked just fine though, CPU usage was down,
although throughput was not so good as with 2.4 (about 10 % overall
degradiation in the "intelligent" read and write, but per-char results were
much better). However, with deadline I got a clear improvement over 2.4 in
the numbers and the behaviour during bonnie++.

These experiences suggests that I until this get solved I'm going to keep
anticipatory scheduler off. Either it's directly responsible or just brings
out some misbehaviour in another place. And yes, I have machines where AS
never appears to cause any particular kind of trouble, so I got no idea
what's really wrong. Other people who have had this kind of issues might
want to try the same trick and report with results.

Antti
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/