Re: performance (was Re: cache not up to date)

At Mon, 24 Jun 1996 20:22:13 +0800, Ong Beng Hui writes:
>> We are having a DEC AlphaServer 4100 with 512 Meg of memory
>> and 50 Gig of disk space. We are increasing the disk used
>> by the proxy slowly to track the approximate memory usage.
>> Currently, we are using 24 Gig and have around 1.2 milions
>> hits per day with around 40 % average hit and 50+ on a
>> "fine" day.
>>
>> For Digital UNIX, the max file descriptor is 4096 and
>> on peak hour, it can hit up to 600 in used and more.
>> We are monitoring the memory usage to prevent any
>> paging, if possible.

Interesting... Currently our production proxy is a 512MB Pentium
machine running NetBSD and CERN httpd with 14GB of disk for the cache,
and doesn't deal well with our load (60,000 dialup users, up to
200--500 simultaneous connections at peak times, an awful lot of
forked httpd's and paging). We ran a brief trial with Squid under IRIX
but ran into several problems that made us back out... firstly the
machine ran out of kernel networking memory (easily fixed) and
secondly it didn't multiplex IO well enough so the latency on serving
even cache hits was too high when there were a lot of simultaneous
connections. We think we've addressed the second problem with a fairly
large patch to use asynchronous IO under Irix, but won't know for sure
until we run another production trial this week.

I'm curious about the amount of core you have in your
AlphaServer... How big an in-core cache are you using, and what sorts
of resident set size do you see when Squid is under load? How about IO
rates to and from disk?

Pete.
Received on Mon Jun 24 1996 - 06:38:59 MDT