|
Yes, during our tests we?ve seen that there are some problems on real
servers.
When the number of requests redirected to each server exceed a certain
threshold, the sustained rate of the system suddenly collapses.
When this happens the CPU utilization of the system is almost 100%, the
real server is overloaded and it seems waiting for any system resources
to free themselves.
We?ve increased the number of per-process file descriptors, the total
number of file descriptors, the ip port range, and so on?but the problem
still remains.
We've found an old Internet-draft: ?Avoiding the TCP_WAIT state at busy
servers? (T.Faber, J.Touch, W. Yue).
It deals about the problem of the accumulation of connections in the
?TCP_WAIT? state and it explains that ?servers that have many TCP
connections in TIME_WAIT state experience performance degradation, and
can collapse?.
In this document they present the results of their tests showing a 50%
improvement of HTTP connection rate.
This is why we are trying to reduce the number of connections in that
state, not only for aesthetics :-)
I?m not sure this is the problem, but it could be.
claudio :-)
-----Messaggio originale-----
Da: mack@xxxxxxxxxxxxxxxxxxx [mailto:mack@xxxxxxxxxxxxxxxxxxx] Per conto
di Joseph Mack
Inviato: martedì 14 ottobre 2003 13.54
A: ilclaudio@xxxxxxxxxxx; LinuxVirtualServer.org users mailing list.
Oggetto: Re: busy servers & time_wait connections
ilclaudio@xxxxxxxxxxx wrote:
>
> Hi,
> I'm using LVS-DR to balance 12 real servers.
> Each server runs a little search web engine.
> When these servers are overloaded there are lots of TCP connections in
the state "TIME_WAIT".
> I know why they are in that state, because I read some documents about
this argument.
> However I'd like to reduce the number of connections in this state.
> That's I want to reduce the time before a connection is closed
definitively.
> I read an old post of Roberto Nibali (about LVS Performance) who tells
that it is possible the tuning of TCP_TIMEWAIT and CLOSE_WAIT
connections.
> But he only mention this possibility.
> Does anybody know how to do this?
Changing the timing is written up in many places, but the tcpip
parameters
are one level below LVS and you really shouldn't mess with them unless
you've
got good reason.
Why do you want to get rid of the TIME_WAITs?
o Are you running out of sockets? then increase them.
o aesthetics ie you don't like seeing a long list of TIME_WAITS?
then go find something else to concern
yourself with and leave to computer to handle the connections, it's
doing just fine.
When you have 10s of 1000s of connections, you're going to have long
lists
of TIME_WAITs
o is there some pathelogical condition you haven't told us about?
Joe
--
Joseph Mack PhD, High Performance Computing & Scientific Visualization
SAIC, Supporting the EPA Research Triangle Park, NC 919-541-0007
Federal Contact - John B. Smith 919-541-1087 - smith.johnb@xxxxxxx
|
