Sent to a Squid developer I just had a discussion with on IRC...
-------- Original Message --------
Subject: Squid: Conversion of disk cache reads into cache misses when
disk bandwidth is saturated
Date: Sun, 30 Apr 2006 14:07:29 +0200
From: Mark Bergsma <mark@...>
Hi. As discussed on IRC, a summary:
We are having some problems efficiently using disk caches in our
Wikimedia Squid cache cluster. When the disk cache is too big, disk
cache reads can easily saturate all of the available disk bandwidth on
all disks. This makes Squid very slow as it keeps queuing up the
requests. Therefore we deploy many Squids without disk caches and have
them memory only, which results in much faster Squids (lower request
times with much shorter distribution tails), but the total hit rate suffers.
It would be neat if Squid could decide to convert a disk cache read to a
cache miss when the relevant cache_dir is overloaded. That would make
Squid more load-tolerant reduces the need for administrators to tune and
readjust disk caches for each server over time.
As Squid already does this for cache *writes*, this should be fairly
easy to implement for reads too. Let me know when you have something,
I'll be happy to test patches on a test server in our cluster. I'll
stick around on IRC too...