Re: 100% CPU : hit a limit ?

From: Willy Tarreau <w#1wt.eu>
Date: Mon, 10 Mar 2008 23:43:25 +0100


Hi Patrick,

On Mon, Mar 10, 2008 at 11:03:00AM +0100, Patrick Viet wrote:
> Hi,
>
> My HAproxy has just reached 100% CPU :
>
> Cpu0 : 0.7%us, 0.7%sy, 0.0%ni, 74.8%id, 0.0%wa, 4.3%hi, 19.5%si, 0.0%st
> Cpu1 : 6.3%us, 90.4%sy, 0.0%ni, 0.0%id, 0.0%wa, 0.0%hi, 3.3%si, 0.0%st
> Mem: 1023848k total, 857328k used, 166520k free, 280708k buffers
> Swap: 1228792k total, 40k used, 1228752k free, 121624k cached
>
> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
> 15019 root 25 0 108m 107m 408 R 100 10.7 46:50.52 haproxy
>
>
> And by the way, this is not shared between the CPUs as you can see.
> ( Linux version is 2.6.20.7 )
>
> Is that the normal behaviour when I hit a performance / capacity limit ?
> It's using up 40mbit/s on a Pentium 4, with many hits, and currently
> 2500 simultaneous sessions.

that's pretty low for 100% CPU. I suspect that you don't have epoll compiled in. Could you please try to run "haproxy -vv" to see how the build options, as well as report the output when starting with the following config :

# cat >test-pollers.cfg << EOF
backend dummy

        balance
EOF
# haproxy -V -f test-pollers.cfg

> The sites are still running kind of OK even if they are slighly slow...

If you're running on poll() or select(), I think that's pretty much expected, so please run the tests above, and rebuild with "make TARGET=linux26".

Regards,
Willy Received on 2008/03/10 23:43

This archive was generated by hypermail 2.2.0 : 2008/03/10 23:45 CET