Re: [Lse-tech] IPC lock patch performance improvement

[<rw...@ea...>]

> I ran the LMbench Pipe and IPC latency test 
> SMP kernel running on a 4-way 500 MHz. 
> average of three runs.

I generally run lmbench 25 times on SMP and sometimes wonder
if I have enough samples to be really confident in small
differences.  Partly it depends on the lmbench test.
For short lived tests like stat/fstat 25 samples is perhaps
more than sufficient.  For the tests that vary more,
I don't know.

Here's Pipe bandwidth from 2.4.19 on a 4-way box (25 runs):
Max	529.76 MB/s
Avg	447.71
Mean	437.94
Min	346.13

There's a 53% difference between min and max.  Someone good
with statistics could probably calculate the variance and
standard deviation and tell what the margin of error is with
my numbers.  So far, I've just been doing an average of a 
bunch of samples, and focusing on the big numbers.  

My web pages need a dynamic overhaul to make the content
more useful.  I can't put it off forever, can I? :)

-- 
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html