Re: haproxy at amazon

On Tuesday 20 of September 2011 02:02:27 Dean Hiller wrote:
> We are running haproxy at amazon and running some load tests and seem to be
> hitting some bottleneck between haproxy and webservers or haproxy itself.
>
> How can you tell when haproxy is maxed out? Will cpu hit 100% or is it
> some other characteristic? our cpu is 4% and I only have 10 webservers
> and 10 clients, and my 10 clients generate about 1000 requests/second each
> on a socket and each one is stateless independent of the other, no session
> is saved at all.

If you configure it correctly (haproxy bound to some specific core, network interrupts to some other core sharing L2 cache with haproxy core), you should see 100% on haproxy core (70% system, 30% user if running in L7 with few acls and rewrites) and around 25% on core servicing network interrupts. In full http tunneling mode you should see both cores saturated ad 100%.

You should check traffic on your haproxy host both ways using tcpdump.

Regards, Brane Received on 2011/09/20 08:59