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

> 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

I believe we have roughly a similar load. Except, I don't
have as many users as you have. I probably some advice from
you on such a huge users base. You should be approaching
the magic number of 65535 soon.

> 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

Our try on a SGI Challenge S running Netscape Proxy V2 break too.
IRIX is a little sensitive on the mbuf parameter, and I believe
you have to apply a patch to enable some tuning of kernel network
parameter.

> 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.

Great, let us know your result. We still have a Challenge S
with 128 Meg of memory laying around.

> 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?

My understand of core is the amount of memory ? nope ?
The system configuration is a 4100 with 512Meg and no cache.
I dunno if it is a good decision, well, I am not the one
making them.

First of all, I am not running Squid, but Harvest Cache v2.
I assume that they should behave quite similarily. The
RSS is around 300Meg to 350 Meg at the moment.
IO rates now is typically 600-1000Kbps on 3 different
file set on a RAID 0 striping configuration.

I believed that kernel tuning play an important part in
most of the production cache since they are under heavy
load most of the time. My experience with UNIX is that
every UNIX with default kernel parameter will break
pretty fast (30 mins ?) when plug into our network as
production machine.

Any tips on Digital UNIX kernel parameters ? anyone ?

*8)
Ong Beng Hui
ongbh@singnet.com.sg
...and they lived happily ever after
Received on Mon Jun 24 1996 - 07:52:21 MDT