pre-forking problem

[Rob Hartill <ha...@ooo.lanl.gov>]

The sys-admin installed 0.7.2h at Cardiff with the default 20
children... it soon ground to a halt, the reason being that Cardiff
gets hit a lot by the yanks, and the connection today into 
JANET (Joint Academic NETwork) inside the UK is as flakey as 
the paint on a Russian airliner.

All the children get swallowed up sooner or later by connections to
slow clients.

Upping the number from 20 to 40 helped, but again, with 60% packet losses
on most of the connections, local users faced unpredicatable delays
on making connections - it could be instant, or it could take 20+ seconds.

The server has been running with 80 children for about six hours now.
The load went as high as 60, but has settled down to somewhere between
5-10 (packet loss now ~20%, with an average 230ms round trip for 64bytes)
(our CGI load keeps the load average ~3 most of the day)

I might lower the TIMEOUT setting to something like 30 seconds. 
(that'll mean the entire request will need to be read in 30s, but
any block of output will get 30s).

Cardiff serves a hell of a lot of CGI, which might be a contributing
factor. However, lots of the CGI is cached, so if a connection breaks,
chances are, the output is sitting around ready for the inevitable retry.

It'd be really useful to find a way to prioritise local connections..
I can't see the locals putting up with these kinds of delays just because
a part of the net that they don't use is unreliable.


rob.