Geez.... such stresstesting!

From: Mr. Corn Hole e Yoh <>
Date: Wed, 22 Apr 1998 18:15:47 +1000

> Date: Mon, 20 Apr 1998 23:48:24 -0700 (PDT) the performance Psycho writes:

> And I found out that my numbers were quite off. At the peak rate,
> The machine handled 343 requests in 1 second (holy sh*t!).

Geez. Are you using -d (buffer log output) and upping the logfile sizes?
You might as well, since at that kind of
connection rate I imagine you'll be issuing lots of logfile writes! :)))

> I'm going to do some more hacks, one of them involving raising the
> listen queue to 1024. Another involving doing the 2.0.33 function
> mods (I've got it maxed at 1024 because (select?) breaks things).

More geez. I thought the listen queue was already pretty big... like 256 or
I imagine the only reason you can actually achieve this is because you're
retrieving a fairly small set of pagedata that the "hot ram cache area" can
serve out of. It'd be intriguing to see the degradation as you start to
rely on the kernel's file caching.

> I've got to compile it with poll instead, and get the hacks so
> the machine'll boot. I figure that the absolute maximum upper
> limit is 1024 hits per second.

I daresay even 1/3 of that for a lowly PPro box would be a very respectable
figure. Anyone who runs large webfarms care to share what their incoming
traffic's like? I know supports 3200+ users suckin' through
ftp, but that's not quite the same thing...

I imagine something DEC or Sun would be the incoming conns/sec leader.... but
a PC shitbox sustaining 300-500 conns/second's not too too shabby.
Received on Wed Apr 22 1998 - 01:18:33 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:39:48 MST