[1372] in Kerberos_V5_Development
"benchmark" numbers not so bad after all
daemon@ATHENA.MIT.EDU (Ken Raeburn)
Wed Jul 3 02:39:13 1996
Date: Wed, 3 Jul 1996 02:38:02 -0400
From: Ken Raeburn <raeburn@cygnus.com>
To: krbdev@MIT.EDU
Cc: kerberos-dev@cygnus.com
Silly me. I was aware that client CPU and network bandwidth were
likely to be a possible bottleneck with this type of benchmark. I
thought I had gotten around them, with the use of multiple client
machines, but I was wrong. (Guess I picked the wrong machines.)
So my "benchmark" numbers have jumped considerably. With the server
code we're running now (replay caches disabled, some of the other
tweaks I mentioned, and not compiled for profiling this time), I'm
getting about 60 TGS_REQs handled per second, averaged over two
minutes. For pure AS_REQs, I get over 80 per second.
The server got up over 90% CPU usage, where before I couldn't get over
60%. So despite the full UDP receive queue I thought I was
maintaining in my earlier tests, CPU power on the server wasn't the
limiting factor -- though I'm getting closer to that point now. And
when I add another reasonably fast client machine, the numbers from
the other clients drop, and the CPU usage doesn't increase
significantly. So *now* I think I may be reaching the limits of this
server machine.
This doesn't necessarily invalidate the profile data I got before; it
just means the resulting performance numbers may have been skewed.
(Probably by limiting the higher numbers slightly, if anything.) I'm
going to do a few new tests without the profiling code anyways.
