Re: [squid-users] Squid 3.3.2 and SMP

From: Alex Rousskov <rousskov_at_measurement-factory.com>
Date: Mon, 11 Mar 2013 12:36:55 -0600

On 03/09/2013 12:24 AM, Amm wrote:

> In short, for best results and to make sure that each worker uses separate core
> and dont end up using same core, one must use cpu_affinity_map as well?
>
> Am I correct?

Based on our experience, yes. Similarly, keeping other significant
activities (e.g., NIC interrupt handling) away from Squid CPU cores may
help with performance. I suspect this "we know better than the kernel!"
approach works well for Squid-dedicated servers. It may not be a good
idea for general-purpose servers where there are many different
processes competing for scarce CPU cycles. Kernel algorithms ought to
work better than rigid affinity schemes in those cases.

BTW, there are command-line tools that can set or change CPU affinity of
a process (e.g., taskset), but cpu_affinity_map is usually easier to use
when it comes to Squid kids.

Cheers,

Alex.
Received on Mon Mar 11 2013 - 18:37:09 MDT

This archive was generated by hypermail 2.2.0 : Tue Mar 12 2013 - 12:00:04 MDT