[httperf] time wait limits

[david at lang.hm]

I spent a bit of time fighting httperf over the weekend, and from reading 
the list archives there's a lot of problems in figuring out how to trigger 
parallel requests during testing.

my test environment

server  Debian Etch running apache 2
32G ram, 4x dual core Opteron CPU's, hardware raid controller with battery 
backed cache and 2x 144G 15K rpm drives

client same

using the apache benchmark tool (ab) I could get 21,000 connections/sec 
with apache prefork, and 23,000 connections/sec with apache worker

with httperf I had trouble getting anywhere near those numbers.

to deal with the problem of running out of client ports I set 
net.ipv4.ip_tw_reuse=1 and net.ipv4.ip_tw_recycle=1

this lets me get up to several thousand connections/sec

by setting the --rate to several thousand and the --num-calls to 20 or so 
I was able to get it to start doing more connections in parallel

the problem I keep running into is the file descripter limits. why is it 
that the max file descripters are capped at FD_SETSIZE no matter what the 
ulimit is set to?

it would be handy if the option were to be added to be able to specify the 
source IP address to use (or to be able to provide a list/range of 
addresses to use)

with modern hardware it's really easy to exceed the limits

David Lang