On 9/17/09 3:18 PM, Willy Tarreau wrote:
Hi Marc,
On Thu, Sep 17, 2009 at 10:49:44AM -0400, Marc wrote:
I didn't specify a TARGET option. This was before the Makefile was changed
to prevent that mistake.
this should not cause an issue either. Can you check if the CPU is
spent in user or system ? User could mean that you have a lot of
ACLs or regex. System could imply other issues (conntrack, poor
network driver, ...). If you observe almost 100% user, it would
mean you encountered a bug causing haproxy to loop at some point,
for instance trying to process an event which has expired, or
something stupid like that. In such a case, the issue would
continue even if you stop the load (eg: if you would unplug the
wire for several seconds).
2800 reqs/s is not much at all. My pentium-M laptop does 4000
(or maybe 6000, I have a doubt now) on low-power mode at 650 MHz.
However, I see that you have an old 2.6.18 kernel, which contains
the old stupid O(1) scheduler as well as strange artefacts in CPU
usage measurement (you could see a sudden 100% wall when reaching
50%, as well as a sinusoidal load between 0 and 100% without
changing the workload at all). So I would not completely rule out
that possibility either.
Willy
RedHat and CentOS have brainwashed everyone into think old Fedora is better
than new Fedora. I just get tired of hearing people parrot the myth that
CentOS is "more stable" than Fedora. CentOS is old Fedora, no secret stability
sauce added in by the CentOS people AFAIK.
