Re: Introduction

On Thu, 2002-11-21 at 07:37, Brian Akins wrote:
> Some more info on what we are doing:
>
> All is reverse proxy.
>
> We have a fairly small object set (10k) the most active can fit into
> 512MB of memory. We have tested squid on a dual 2.4Ghz P4 with 2 GB ram
> and dual 18GB 15k drives. We are running two squid instances per box.
> I know this isn't optimal, but it's what we have.

Actually, it's pretty good. Squid doesn't scale m-cpu wise in any
significant fashion today, so 1 instance per cpu is wise.

> What we are seeing is that as we approach 175k requests/min/instance,
> response time begins to climb and then the server becomes unresponsive.

Yah. Squid-3 has more work from Adrian on the comms interfaces, which is
starting to pay off. poll() is a huge killer for squid. Approaches such
as the rt-signals patch, kqueue, or perhaps someone porting squid to use
epoll() should all help with this.

> I grabbed the latest 3 snapshot and it seems faster, but I couldn't get
> the CPU profiling to compile, so I'm not sure were I should begin :(

What occured with the profiling code? Do you have an error? What build
options?

Rob