We have a config of 2 gs140s running Tru64 v 4.0F.
"frontend" runs squid and apache, and serves static content. Request for
dynamic content are sent to "backend", which runs apache and locally
written apps.
Users are reporting the config to be very slow at times, so I investigated
"backend" during a slowdown. Network traffic overall on "backend"
appears to be around 500-1300 packets/second overall. This seems
divided almost 50/50 between traffic to 2 database backend hosts
(gs140s) and traffic to/from "frontend"
"backend" is only using about 1/4 of its memory, and is about 56% idle.
The one striking thing I see is the nature of the traffic to "frontend".
At any given time when these slowdowns occur I see on the order of
1,800 network connections to "frontend", of which about 1,775 are in
a TIME_WAIT state! Also, the turnover is continuous - snapshots of connections
taken about 5 seconds apart showing "ESTABLISHED" connections to "frontend"
showed completely different connections.
Connection btwn "backend" and "frontend" is over 100BaseT half-duplex.
I'm wondering if tuning of tcp parameters is in order here.
We've done some tuning already, based on a Compaq document for "busy
internet servers". Parameters I've tuned include:
inet:
ipport_userreserved = 65000
ipqs = 16
tcbhashnum = 16
tcbhashsize = 16384
tcp_keepalive_default = 1
tcp_keepidle = 1200
Anyone have any particular insights into this kind of behavior?
The only other thing that comes to mind is perhaps apache is not
efficiently handling the number of connections. There were 109 httpd
processes on "backend" a moment ago, but that is declining
right now, currently about 72.
Huge TIA!!!
Judith Reed
jreed_at_appliedtheory.com
Received on Tue Feb 20 2001 - 14:51:17 NZDT