Performance of freeBSD 5.1

[TeJun Huh]

 Hello,

 I've been conducting performance tests using simple HTTP proxies and
web polygraph <http://www.web-polygraph.org/>.  Web polygraph is
believed to generate realistic Internet load (high number of sessions,
many idle, req rate independent of resp rate, and so on).  I built
several equivalent HTTP proxies using various concurrency mechanisms
and comparing those on linux and freeBSD.

 The following result is obtained on p4 xeon 2.4g (UP kernels).
Custom version of Kqueue is used on Linux.  FreeBSD 5.1 is updated to
the latest cvs tree three weeks ago (due to thread stability issues)
and compiled with GENERIC option.

	FreeBSD				Linux
	Select		Kqueue		Select		Kqueue
rps	400		500		500		2500

 I feel something is very wrong with FreeBSD kqueue result.  Used
tuning parameters are

* Loader tunables

kern.maxusers="512"
kern.ipc.maxsockets="200000"
kern.ipc.nmbclusters="51200"
kern.ipc.nmbufs="102400"
net.inet.tcp.tcbhashsize="65536"

* Sysctls

sysctl -w net.inet.ip.portrange.first=1024
sysctl -w net.inet.ip.portrange.last=65535
sysctl -w kern.ipc.somaxconn=1024
sysctl -w kern.maxfiles=200000
sysctl -w kern.maxfilesperproc=200000
sysctl -w kern.threads.max_threads_per_proc=100000
sysctl -w kern.threads.max_groups_per_proc=8000

 As soon as the test machine becomes available, I'm gonna run tests on
freeBSD 4.8 and compare the results.  Any suggestions are welcomed.

P.S. Please don't forget to CC me. TIA.

-- 
tejun