> I'm trying to do a rough benchmark comparison of Apache 2.0 with your patches
> and "Proprietary Web Server X", both running on an 8-processor SGI box.
>
> Any suggestions? I've been running a bunch of perl scripts on people's
> desktop boxes, each doing a fetch of a 40k page of static HTML. The two
> servers look comparable, but 1) the times are all over the place, seeming to
> imply that I'm actually measuring overhead from the desktop machines and not
> the server response time, and 2) the cpu utilization on the server never
> breaks 2 or 3 percent, for either web server.
I don't know how to share my years of experience benchmarking web
servers in an e-mail message, but I can sum up a few shiny bits of lore
and some starting points. First, if your server's CPUs aren't 100%
utilized you're benchmarking your clients or the network not the web
server and your comparisons are meaningless. When I run benchmarks I
need two to four load generating clients per server CPU over a fast
network like a private 100BaseT or gigabit ethernet. For instance, to
max out a 4 CPU Origin2000 I use 16 2-CPU Origin200s connected via 4
private GbE networks. Double all that for 8 server CPUs. Anything less
just won't place enough load on the server.
Requesting a single web page over and over does not accurately model
real-world loads on web servers. The SPECweb96 benchmark requested a
large number of static files over and over and is now obsolete due to
its limited capabilities. SPECweb99 issues both static and dynamic
requests. Read more about them at http://www.spec.org .
There are lots of performance analysis tools available for your
unidentified SGI box. Check out SpeedShop, osview, perfex, sar, par,
Performance Co-Pilot, xbstat, and WindView for starters. Information
about all of these tools is available from http://techpubs.sgi.com .
Remember to tune all parts of your system: the web server application,
the kernel, device drivers, the hardware.
Read SGI's Web Performance Tuning Guide at http://www.sgi.com/tech/web/
and SGI's SPECweb96 disclosures from www.spec.org, such as
http://www.spec.org/osg/web96/results/res2000q1/web96-20000124-03351.html
for descriptions of how to tune your system for high performance web
serving.
Since you decline to identify your "Proprietary Web Server X" I can't
quote a specific performance comparison, but I can say that the
Accelerating Apache Project's patches make a properly tuned
Apache/1.3.12 run faster than every other web server I tested on Irix
and Linux, including the latest from Zeus and Netscape/iPlanet, on the
SPECweb96 benchmark. I'm tuning Apache/2.0 now, focusing on SPECweb99,
and I expect it to perform similarly by the time I finish.
Finally, since you're measuring the performance of an SGI machine, SGI
may be able to help you. Contact your local SGI service representative
or SGI Managed Services and tell them I sent you.
http://www.sgi.com/services/managed_services/
Have fun!
--
Michael J. Abbott mja@xxxxxxx http://reality.sgi.com/mja/
|