Hi,

i am currently trying to setup pound for load sharing between different 
servers. Main purpose of the intended config should be to distribute
file-downloads to itīs own backend using thttpd server while the normal
requests go to an apache 2 server. There is one server for pound being
connected to the external world via one ethernet line, and to the other
server via its own private range ip subnet on another interface.

Distribution of requests seems to work as expected, however i get 
various requests from users complaining about broken or timed out 
downloads. So far i was able to track it to be very slow connections
below 10kb/s. I am not sure yet what happens exactly but suspect
some timeout of the connection between backend and pound because
the preread part of the rather large requests (about 15 to 30 MB)
is still being delivered on the other side. When i access the
same server without pound all the downloads work, so i guess
the problem must be somewhere in the combination with thttpd.

Can anyone hint what parameters and tcp/ip sysctl values i should
have a look for? All boxes are running a 2.6 kernel based linux
system.

Thanks,

Markus Kramer

Synalogic e.K.
Im Hirschfeld 28
52222 Stolberg
GERMANY

On Friday 14 January 2005 22:16, Markus Kramer wrote:[...]

If your servers are really slow you may indeed see timeouts from Pound. Try 
playing with the Server parameter and see if that helps.

Please note that on VERY long timeouts the client browser itself will also 
time-out (I think common browsers do that at about two minutes).[...]

Hi Robert,

thanks for your quick reply. However the problem is not the server 
being slow.
That one actually delivers at full rate or at least about 50 to 60 
Mbit/s to pound
via internal network till the buffers are saturated. However in fact 
the client
is the slow end only fetching that data at 3kbit/s. Everything runs ok 
for quite
a while and pound delivers a full rate 3kbit/s stream to the client, 
and then
it suddenly just closes down the connection. Up to now i was only able
to do a tcpdump for the outgoing stream to the downloading client, and
from that one i can see that pound closes down the connection not my
testclient.  I will try to catch a log of the backend connection as 
well to
see which end for that is actually closing down the link.

Regards,

Markus Kramer
Synalogic e.K.



Am 15.01.2005 um 11:56 schrieb Robert Segall:
[...][...][...]

On Saturday 15 January 2005 12:20, Markus Kramer wrote:[...]

This disparity probably implies long waiting times for the back-end. It may 
well be that your back-end server decides to close the connection on its own. 
Check your logs and compare them to the Pound logs.[...]