Re: [squid-users] What load can SQUID cope with in accelerator mode?

From: Duane Wessels <wessels@dont-contact.us>
Date: Sun, 18 Mar 2001 20:34:45 -0700 (MST)

On Thu, 15 Mar 2001, Ole Moller wrote:

> >I've done tests with polygraph and achieved 500+ requests per second
> >with similar workload characteristics (90% DHR, 10K file size).
>
> Would you guys please share some hints about how you accomplished that?
>
> I have a very similar system (athlon-800, 1.25Gb ram, 4x4gb scsi disk). It
> is running two apache webservers one with all the fancy stuff and one to
> serve static files (mostly zip-files and a couple of png's). At peak I have
> 1500 concurrent http-demons of the later running and up to 30-60 seconds of
> queue for getting a request served. My best guess is that it was
> approximately 20 req/s at peak. In the static-serving http I have a little
> perl-script running that translated the request to a physical file via
> mod_rewrite.
>
> Then I got the idea of putting a squid up as a accelerator, moving the http
> demon to 127.0.0.1, using the same perlscript as redirector. At first
> glance it worked like a charm. When I was the only user I got almost full
> throughput when requesting a uncached file and it was even faster when the
> file was cached. Then I onlined it, and cpuload, memory usage and such
> dropped dramatically, the number of http-demons was reduced to about 300 -
> fine. Then I tried it for my self, I had to wait about 3 minutes before
> squid started to serve my request and when it did I got about 2kb/sec
> (instead of expected 1mb/sec). When squid had run for about 10 minutes it
> killed itself with a message of "FATAL: Too many queued redirector requests".
>
> I believe that this not is the real cause of death considering that I had
> started 30 and one did the job fine when running from apache. Have you any
> idea of what could be blocking? disk io?
>
> some facts: OS is FreeBSD 3.3 (highly tweaked, filedescriptors and such are
> configured for high load)

You can find some details at
http://www.squid-cache.org/Benchmarking/Surrogate07/

I wasn't using Apache. In my test, Polygraph is serving responses.

Also I wasn't using redirectors. That seems like a likely bottleneck,
especially with just one CPU.

You may want to grab my "squid-rrd" stuff from
http://www.squid-cache.org/~wessels/squid-rrd/. It may help
you see if your bottleneck is CPU, disk I/O, etc. Unfortunately
those scripts don't grab any information relating to redirectors, but
its probably not too hard to add that.

I lied about achieving "500+" TPS. It was closer to 480.

PS, I could never get FreeBSD 3.x to stay running very long with
Squid. It always ran out of vnodes, locking up the system. I've had
much better luck with 4.x.
Received on Sun Mar 18 2001 - 20:34:46 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:58:43 MST