Squid optimization guide from last.fm of all people

Benchmarks don't backup many of these claims. Search "joe@swelltech.com squid performance" for my authority to make such a claim in the face of an extremely well-written blog post from a very smart fellow.

But the core bits of ReiserFS, raised file descriptors, and balancing RAM to disk are valid.

The aufs threads just aren't useful over about 24 even on the fastest machine--they are disk limited, not CPU bound. Squid is not multi-threaded, but there is a disk IO thread pool option (the aufs bit in the article). Going crazy with threads is counter productive and just eats memory.

Most of the configuration and proc options covered are just superstitious from a performance standpoint--they're generally fine at their defaults, and it's usually more productive to figure out what is binding performance and then turn the knobs with that knowledge rather than blindly cranking everything to eleven.

So, not a bad article, per se...but one that is rife with superstition. Squid brings it out of people like no other project I've ever seen (aside from maybe filesystems).