client keepalive on high trafic site

Václav Bílek v.bilek at
Tue Sep 8 12:37:23 CEST 2009

Poul-Henning Kamp napsal(a):
> In message <4AA620D7.1060807 at>, =?UTF-8?B?VsOhY2xhdiBCw61sZWs=?= writes:
>> our sess_timeout is set to 1sec alredy :(
> Then I doubt keepalives are your problem
> It is much more likely you have workerthreads waiting to push stuff
> to slow clients.
I think that is not the problem... when I disable keepalive by seting
"connection: close" header then i get 4x higher throughput ( but works
bad in IE6)
... till the 8K established connection everithing works fine (low
latency responses, only 10%cpu usage on server, load around 1-2)

...  the problem is that ther is by estimation around 150K of clients
which is imposible to handle with keepalive enabled

the webserver cluster we use till now had the same problem ... keepalive
even set to 1 second causes so many opend connection that system stoped

> Consider increasing your kernels tcp sendbuffers ?

set  large enough for 95% of requests

is it at all posible to implement global sesion_timeout=0? or is it a
design problem?

More information about the varnish-misc mailing list