Re: Squid 2.x performance curioisities

> Oskar Pearson wrote:
>
> > Dancer: I know you don't like kernel 2.2, but give it a try. At least
> > then you can narrow things down, ok?
>
> Linux 2.2 also seems to have quite a bit faster disk I/O system than
> Linux 2.0, perhaps a result of better buffer management..

Well, we're running Linux 2.0/Squid 2.2 with async-io and I've had to increase
NUM_THREADS to 128, and it even gives me warnings now due to backlogged deletes.
However, even with NUM_THREADS at 16, it runs very well and the warnings are
little more than something to try and stop being logged :-)

That's with 4 x 9Gb Seagate UltraWide Barracuda drives; I think it's mainly
due to the unlink() time in the 2.0 kernel.

Of course not using async-io would be slow as you go back to the performance
nightmare of 1.1.x which we liked so much we stuck with it to further the
impact of delay pools on the students :-)

Oh, and we're finally moving from 1.1.15 DELAY_HACK to 2.2 DELAY_POOLS for
our own use for our "slow student proxy" (previously I'd just been developing
delay pools in 2.x as a potential future change for here). Using that combined
with the "myip" ACLs and a small hack to log the "myip" value in the unused "-"
in common log format to basically run everything off one cache as 'virtual
caches' with different ACLs, delay pools and so on. The great news is that a
class 3 delay pool and some fairly complex ACLs don't seem to have any
significant CPU cost - the CPU load at present is not noticeably different from
when it was running exactly the same software compiled without delay pools,
based on the MTRG graphs.

David.
Received on Tue Jul 29 2003 - 13:15:58 MDT