Re: squid-2.3-200007042300 and AIO_PROPER_MUTEX

Wayne Smith wrote:
>
> Currently running Squid 2.2 Stable 5 with WCCP back port and HNO patches
> doing async-io. Seems stable, and can handle more requests/sec than 2.3
> Stable 3 without async-io.

Thanks.

> 2.3 Stable 3 with async-io enabled doesn't seem all that stable. With
> squid-2.3-200007042300 patched applied to Squid 2.3 Stable 3, will aync-io
> know be 'happier'? Will the AIO_PROPER_MUTEX make life happier?

async-io in Squid-2.3 isn't stable when put under load. Period. No time
or plans to fix it for the 2.3 release. async-io in 2.3 is behaving a
lot worse than unpatched async-io in 2.2, and a unpatched 2.2 does not
behave well at all.

The situation is hopefully a lot better in the 2.4 releases which should
not be very far from a "STABLE" release now. There are some minor issues
with the 2.4 version still left on the todo, and it needs a fair amount
of testing before 2.4.STABLE is released, there is no known major issues
I think. (note: The current 2.4 squid-head version is wastly different
from the 2.4.DEVEL2. There is no point in even looking at the DEVEL2
release nowdays)

> I like the added functionality of Squid 2.3 Stable 3 where you can set a
> limit on minimum object size, the recent patch for size 0 (frontpage anyone?
> ;) and LFUDA storage replacement policy. Would also like to have the webdav
> methods added. But because of the load on the box, I need async-io. The
> box is beefy (640 Megs, two 17 gig SCSI-2 ultrawides), but without the
> async-io, can't keep up with the number of requests per second. Can't wait
> for 2.4 and DISKD! =]

My 2.2 patchset includes request content-length:0 fix, most WebDAV
requests (at least all which are present in 2.3.STABLE3).

> Henrik, is async-io on 2.3 Stable 3 reliable after squid-2.3-200007042300?

See above.

> And because the developers probably don't hear it enough, thanks guys!
> Pretty impressive work!

Thanks.

--
Henrik Nordstrom
Squid hacker