|
On Thu, 2008-02-28 at 17:02 +0100, Rai wrote:
> I'm running a load balancer with ipvsadm with 10 lighttpd servers.
> Serving more than 6 million pages per day.
How many individual object fetches does that equate to? What are the
mean, median max an min object request rates? I'm not being nosey - this
could be useful info.
> Sometimes i have 3 (or 9) seconds of delay load balancer is serving pages.
Sounds like a TCP retransmit problem to me, indicating packet loss
somewhere in the path. If I recall correctly, an unacknowledged SYN will
be retransmitted 3 seconds after the first one, then there's an
increasing delay such that:
Packet Time
1 0
2 3
3 9 (3+6)
4 21 (9+12)
5 45 (21+24)
...and so on, until the timer elapses and the connection fails
completely.
> If i request a page directly to the lighttpd server, i haven't any problem.
When you test, are you testing an unloaded system? By this I mean: you
test a single webserver which is *not* serving any other requests; you
also test a single, unloaded load balancer which has the ten machines
behind it but is *not* serving any other requests at the same time.
Alternatively, are you hitting in-service machines here? If you are,
your packet flow through the director is (in theory at least) 10 times
that of an individual web server.
Somewhere you're dropping packets. You may want to make sure your
director isn't overloaded or has a NIC or switchport which is reporting
errors.
Graeme
|
