recommended production environment

[henpa]

I've been doing many benchmarks these last few days using different options for environments. I am testing this with a few vmware linux boxes on my Windows desktop. I know this doens't represent a production server, but I believe it does give me some indications about best possible performance configuration, for later deployment on a real hardware.

All benchmarks are done using "ab -c 10 -n 10000 http://localhost/hello" with a simple perl fcgi script. Both boxes have identical lsws std-2.1.12 configuration.

These are the best benchmarks I can possibly get:

debian sarge 3.1 (perl 5.8.4 installed)
with kernel 2.4.27 ~ 1500 reqs/seg
with kernel 2.6.8 ~ 700 reqs/seg

arch linux 0.7.1 (perl 5.8.8 installed)
with kernel 2.4.32 ~ 1150 reqs/seg
with kernel 2.6.15 ~ 600 reqs/seg

I have repeated these tests many, many times. And they tell me that LSWS performs much better on kernel 2.4 than kernel 2.6.

So, my questions are:

1) Does it make any sense of LSWS running slower on kernel 2.6? Can others confirm this? (I am running the ab benchmark util and lsws on the same linux server, would this cause lower performance hit on kernel 2.6?)

2) What is the lab environment on litespeedtech? Can I assume "Fedore Core 3 with kernel 2.6.10-1.760" as noted on the benchmark pages? If you tried a 2.4 kernel, would the benchmarks get higher?

3) Would glibc libraries, perl version and other packages also influence the benchmarks? If so, does any one have best version recommendations?


Thanks, Henrique.

 

[mistwang]

It is hard to tell the real performance under vmware.
Our test shows that there is no big difference between linux kernel 2.4 & 2.6.

But we do find the benchmark result is much worse when ab in apache 2.0.x is used. So always use an "ab" from apache 1.3.x.

And we found linxu 2.6's poll implementation is not as good as that in 2.4, especially when you run the test tool in the same box. So for 2.6 kernel, "epoll" should be used instead of "poll".

 

[henpa]

George, thanks for your reply. I changed it to "epoll", but I didn't notice much change, at least with my two linux boxes inside vmware. Maybe kernel 2.6 doesn't perform well with vmware drivers... Unfortunately, I don't have a "real" server available for testing at the moment.

I would love to see other people's experience with other distribution and kernel comparisons.


Thanks a lot for your help.