On Wed, Feb 11, 2009 at 10:52:42PM +0000, Robert Simmons wrote:
> Hello Willy,
>
> Thanks for your quick reply. I've just re-tested this under 1.3.15.7
> (on my other box) and using the command below it does compile
> successfully, so it would seem to only affect 1.2.18 - which doesnt
> seem to have the file structure you mentioned, I dont see a compat.h
> anywhere in the tarball.
No there aren't any in 1.2.18, indeed. Maybe in 1.3 the problem was
incidently solved by the addition of another header before netfilter.
> I need to check whether using the compiled version solves the
> saturation issues I've had whilst running tests, with the CPU sitting
> at 100% and no further connections being accepted.
there are two cases where this can happen :
- netfilter+conntrack loaded with default config on the machine
- frontend's maxconn too low : the poller calls event_accept()
which cannot do anything then returns. But the accept socket
should almost immediately be disabled by maintain_proxies(),
so it should not be called again.
- global maxconn too low. The poller calls event_accept() which
accepts then drops the connection. Maybe your tester tries to
immediately reconnect in this case, leading to maintained
saturation and the problem you observe.
Please try to push global maxconn and your frontend's maxconn to
very large values.
> I would be
> interested however in discussing my project with you off-list, I'm
> interested in hearing your thoughts on my theory (apologies for the
> obfuscation, but its not a protected idea currently).
OK if you want. But I prefer to tell you now that I often take a lot
of time before replying to long comments on ideas, simply because
the response itself takes a lot of time and concentration that I
rarely find.
Regards,
Willy
