Re: [SQU] multiprocessor machines

From: Colin Campbell <sgcccdc@dont-contact.us>
Date: Wed, 3 Jan 2001 08:21:39 +1000 (EST)

Hi,

Are we not adding a lot of complexity to squid to take advantage of
SMP? Splitting threads across multiple processors can slow things down if
you startrelying on mutexes or sharing data because of (processor) cache
coherency problems requiring memory accesses all the time, flushing the
cache etc.

Why not something simpler like adding SO_REUSEADDR to the socket creation,
running multiple squids (eg one per cpu), splitting the cache between the
squid instances (eg with 2 cpus and 4 disks have squid0 on cpu0 using
disks 0,1 and squid1 on cpu1 using disks 2,3) and treating them all as
peers?

Has anyone tried something like this?

Taking this one step further would require the squids being able to
operate on one cache. I haven't deepe enough knowledge of squid's
internals to know how feasible this would be. Comments?

Colin

--
To unsubscribe, see http://www.squid-cache.org/mailing-lists.html
Received on Tue Jan 02 2001 - 15:24:01 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:57:19 MST