Re: [SQU] current CVS ... diskd config args changed ...

Chris Conn wrote:

> Joe Cooper wrote:
>
>> I benchmarked them a couple of months back. aufs was, then, about 10%
>> faster than diskd on a dual disk system.
>>
>> I have not yet benchmarked a recent 2.4PRE-STABLE Squid, however.
>>
>
>
> Hello,
>
> I have read the following URL following your postings in this thread,
> which have sparked my interest:
>
> http://www.squid-cache.org/mail-archive/squid-users/200101/0798.html
>
> Is it inferred with 2.4-PRE-STABLE that we can once again use ASYNCIO on
> Linux GLIBC2.x systems? Would it be acceptable to use ASYNCIO and aufs
> on such a system? With PTHREADS?
>
> To be honest, I have tried 2.4PRE-STABLE with a Linux system, PIII
> 800Mhz, 2x15GB partitions, 768MB RAM and a 80M/sec SCSI card and the
> system's load average under 2.3STABLE4 is about 0.80 at 2000 req/min,
> and using diskd with 2.4PRE-STABLE on the same system, my load average
> would hover between 2.5 to 3.0, so I obviously reverted back to
> 2.3STABLE4. My system's library is GLIBC 2.1.3.
>
> Also, there is a bug mentioned in the bug watcher about delay pools.
> What is the status on this? Is the patch mentioned functional? I am
> referring to http://www.squid-cache.org/bugs/show_bug.cgi?id=84 .
>
> Thanks,
>
> Chris

2.3 has significant problems when using aufs under load. I found that
it drops hits at an alarming rate. (Hit ratio begins to fall off at
about half of system req rate capacity, and drops from there...leading
to only about 25% hit rates where there should be ~50%.) I also ran
into numerous stability problems under very high loads. So I gave up on
ever using it, and have performed no further testing (we're skipping
straight from 2.2STABLE5+hno+WCCP+some other backported stuff, onto 2.4).

2.4 is faster than 2.3. But only in the sense that it can be used at
higher loads than 2.3 without losing significantly on the hit ratio.
2.3 was way too aggressive in shedding load by dropping swaps to disk
out of the queue...thus it will be easier on your processor and disk,
but will show very poor hit ratio if really loaded. So 2.3 and 2.4 will
actually run successfully at about the same request rates, but 2.4 will
show better hit ratios at those rates. (At least...that's what I found
when I was testing these things. Maybe something has changed for the
worse, or the better.)

2000 req/minute is only about a fifth of what your box will handle if
tuned properly, and a third even if not tuned at all...so you aren't
seeing the hit ratio problem in 2.3. And you really don't need to
concern yourself overly much with performance.

That being said...2.4 hasn't been declared stable yet. There may yet be
some zoomy bits that make CPU usage climb in ways they shouldn't. I
haven't tested it lately to know. I will try to get around to doing
some more testing soon (we just relocated so some of our test machines
are still in boxes).

I know nothing of the delay pools bug. But it wasn't mentioned recently
in discussions on the dev list regarding rolling another PRE version.
                                   --
                      Joe Cooper <joe@swelltech.com>
                  Affordable Web Caching Proxy Appliances
                         http://www.swelltech.com

--
To unsubscribe, see http://www.squid-cache.org/mailing-lists.html