On Wednesday 06 December 2006 16:43, Ahmed Kamal wrote:
Due to the web server hit faced with FC6 release, some actions are being
taken to minimize the chance of facing such issues again. One of the steps
is to stress test our web-server infrastructure to measure our current and
future capabilities. I'd like to run some tests on fp.o web-server, please
let me know your thoughts/comments. Here are some details.
1- Measure our bare (no caching) maximum serving rate
2- Measure our cached serving rate, to assess the implemented caching
3- Gather numbers like (When do we get CPU saturated, RAM requirements ..).
Possibly draw graphs (everyone thinks graphs are cool), the numbers should
help us base future calculations on a solid basis
4- Future: Possibly implement a mechanism to cap the maximum connected
clients to a specific server, to the maximum it can handle gracefully, to
avoid killing a server
I think this is great. I think to get things done right we
will need to do it
in a distributed manner.
1- A script was written which uses apache's ab tool to stress the server.
Script will run on the web-server host.
Is that a fair test? i tired the script on
a box i have which while very
different to the boxes in use i could not get it to break a sweat.
admittedly i was only serving the default welcome page. I will try it again
with a wiki setup and see how it goes then
2- The script fires a total number of connections equal to ten times
maximum concurrency rate (to get good average, and avoid transients)
3- The concurrency rate is sweeped between 10 and 400 (my 1G-RAM machine
swaps at about 100 connections)? any suggestions?
i had up to 2000 connections in
an effort to get thinks choked up and did it
in steps of 100
4- All ab output is recorded for future analysis
I had some
that did not get captured for some reason (1900 and 2000) also i
was left with alot of top processes running
5- A monitoring thread is fired before ab is launched. The monitoring
"top" to record load/cpu/ram/process information in log files as well
6- Tests are repeated with "ab -k" for enabling the HTTP keep-alive option.
Not sure if this is needed, or if it will make much difference! comments?
7- Tests are done once with caching enabled and one more time without
Please let me know your thoughts about the testing setup, should we be
recording more data? should we be stressing the server in a different way,
should we be testing some apache config options ... etc ?
,-._|\ Dennis Gilmore, RHCE
/ \ Proud Australian
\_.--._/ | Aurora | Fedora |