httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dan Kegel <d...@alumni.caltech.edu>
Subject Re: Mindcraft, part II - anyone want to help?
Date Sun, 09 May 1999 15:04:25 GMT
Manoj Kasichainula wrote:
> 
> On Wed, May 05, 1999 at 03:04:49AM +0000, Dan Kegel wrote:
> > In a benchmark like Mindcraft's, should Apache's MaxClients be set to
> > the physical number of client threads, or to a higher value?
> > Don't the server daemons sit around doing a lingering close
> > for some time after finishing each request?
> 
> I'm not the best person to answer this question, but I think the
> server will only sit on the socket if no data arrives for 2 seconds
> (because of old browser behavior), or when the client closes its
> connection.

I doubt the server sits on the socket when the
client closes its connection after it receives the FIN.
Otherwise we'd see massive stavation.

I just heard that there was a bug in the Linux 2.2 kernel,
fixed in 2.2.7, which caused a 200ms delay on connection startup
and shutdown.   Mindcraft observed a peak of 1000 connections/second.
A 200ms delay would result in 200 or so server processes tied up,
unable to do any work.  MaxClients was set to 288, and performance 
went to pot with >160 active clients.  Maybe 128 server processes
were tied up in the 200ms delay?

It would have been very interesting to see what would have
happened if Mindcraft had experimented with raising MaxClients
past 288.  It *shouldn't* make a difference, but then, the
TCP stack shouldn't have that 200ms delay, either.

- Dan

http://www.kegel.com/mindcraft_redux.html

Mime
View raw message