Re: squid-users-digest Digest V98 #40

In message <199801290443.UAA21650@nlanr.net>, squid-users-digest-request@nlanr.
net writes:
 
> Date: Wed, 28 Jan 1998 09:10:05 +0000
> CC: Abha Ahuja <ahuja@merit.edu>
> Subject: Re: OS survey
>
> > I was curious to hear from other squid-users about their hardware
> > and OS recommendations. We are currently running a Linux box (RedHat
> > version 4.1, kernel upgraded to 2.0.33, Squid 1.1.20) And every so often,
> > the squid process crashes for no apparent reason (we think it might be a
> > possible kernel bug)... Also, we are doing some testing on a FreeBSD box
> > as well.

We run a number of squids on linux. The largest being a 40 gig squid,
handling around 180,000 TCP requests/hour in the evenings. This is on
a P-II 266, 512meg ECC ram, linux 2.0.33, redhat, 100 meg full-duplex
ethernet card, handling about 3400 packets/sec of thru traffic.

Squid processes crashing 'for no apparent reason' is almost ALWAYS a
ram problem. The odds against it being a kernel bug are fairly
enourmous. We run a large number of machines on linux, and we run them
very very hard. Normal uptimes on machine with ECC ram are around 3
months (we rarely get higher than that; they get taken down to add new
hardware and/or upgrade existing hardware). Kernel crashes on machine
with ECC ram are non-existant. It just doesn't happen.

Spend the extra money and get ECC ram for your boxes.

Oh, and don't go saying "but win95 runs on it without crashing!". :)
The issue here is L1 and L2 cache locality. While the cache hit rate
is very high, you won't see any problems. (the ram gets plenty of time
to do refreshs, nothing is stealing cycles etc etc). When the
cache hit rate starts dropping and the memory bus is staying 100%
utilized for seconds at a time, _that's_ when you'll start seeing
problems.

Gcc, squid, and anything else with a poor cache locality will tend to
crash frequently on boxes with bad ram, more so than other programs,
which tend to have smaller working sets, and thus better cache
locality, and thus have significantly smaller demands on the ram.

Michael.
Received on Wed Jan 28 1998 - 22:46:31 MST