|
Kees Hoekzema wrote:
>
> On Sunday 06 February 2005 14:03, Joseph Mack wrote:
> >
> > See Sect 5.16, 5.17 and 5.8 in the HOWTO
>
> I've read the howto, at least concerning those points,
OK just checking. I didn't see that you were using LVS-NAT, so there's
no arp caching involved.
I assume in your webpage, that the first graph with a Y axis of 3000
is the response time for the LVS, while the next two with Y axis of 80
are the individual realservers being accessed directly.
> I've already ordered some new gbit nics for the loadbalancer, maybe that'll
> help.
You would hope that this is not the problem (ie the switch handles 100Mbps
same way as it does 1Gbps), but it's one place to look.
There was a race in the Linux 2.2 tcpip stack that gave high latencies
under high load, that caused problems with beowulf setups. The race was
never solved, and the problem was handled by using a shorter timeout
in the 2.4 tcpip stack. Your graphs look like the ones that people got
with beowulfs.
The only way to find out what's happening is to look through tcpdump
logs for the connections that have high latency and to see what's going
on.
Joe
--
Joseph Mack PhD, High Performance Computing & Scientific Visualization
LMIT, Supporting the EPA Research Triangle Park, NC 919-541-0007
Federal Contact - John B. Smith 919-541-1087 - smith.johnb@xxxxxxx
|
