The results of these tests indicate that there may be some odd SMP locking
issues still in 2.5.xx which were either much less obvious or not present
in 2.4.
The oddness I see
As these results show, best results were from using a uni kernel, but an
SMP kernel with "nosmp" was faster than running SMP. The problem is that
with 2.5 kernels the SMP results for multiple runs vary by up to 2:1, and
for uni kernels it's more like 5-10% max. I have tried longer runs, more
runs, and gotten the same results even in single user mode.
The ctxbench measures the number of times two processes can take turns
running using various forms of IPC. I'll post a more complete sescription
of the benchmark and a link later today. The IPC in question is signals,
semiphores, and message queues. Pipes are disabled (didn't work in 2.5.50)
and sockets are not currently tested.
I am going to enable the pipe testing code, and will rerun the tests and
post the results as well as the link to source code by tomorrow at the
latest.
-- bill davidsen, CTO TMR Associates, Inc <davidsen@tmr.com> Having the feature freeze for Linux 2.5 on Hallow'een is appropriate, since using 2.5 kernels includes a lot of things jumping out of dark corners to scare you.
This archive was generated by hypermail 2b29 : Sun Dec 15 2002 - 22:00:13 EST