VM tests on 2.4.13-pre5aa1

[rwhron@earthlink.net]

Kernel: 2.4.13-pre5aa1

I discovered something important for the test results I've
been reporting.  The mp3's that I've been listening to were
not all sampled at the same rate.  That means some of the
comparisons are suspect. 

The mp3's were sampled between 88k and 192k.  I did not notice
the sample rate affecting whether an mp3 skips or not.  
I.E. an 88k mp3 and a 192k mp3 skip about the same on a
kernel/test that sputters.  There probably is a difference, 
but it isn't obvious.  So the subjective reports on sound quality 
are reasonable.  In the future, I'll make sure comparisons that
include timing are done with comparable mp3's.


Timing variance:

mmap01  Low time: 4:13  High  4:29
mtest01	Low time:  :43  High  1:10

I'm not saying the difference between the high and low times
is the variance between an 88k and 192k mp3. For mtest01 it 
could be, because that test is short enough to run a couple
times during one song.  mmap01 may have had part at 88k and part
at 192k.

Okay, with the disclaimers out of the way.  Here are the results:


mmap001

Average for 5 mmap001 runs
bytes allocated:                    2048000000
User time (seconds):                19.172
System time (seconds):              15.182
Elapsed (wall clock seconds) time:  258.82
Percent of CPU this job got:        12.80
Major (requiring I/O) page faults:  500169.0
Minor (reclaiming a frame) faults:  32.0

mtest01 

Averages for 10 mtest01 runs
bytes allocated:                    1251370598
User time (seconds):                2.079
System time (seconds):              2.849
Elapsed (wall clock) time:          54.075
Percent of CPU this job got:        8.70
Major (requiring I/O) page faults:  107.2
Minor (reclaiming a frame) faults:  306293.2


Even though I made the disclaimers, I will note that 2.4.13-pre3aa1
was doing the mmap001 test consistently (4 runs) at around 210
seconds average.  2.4.13-pre5 and 2.4.13-pre5aa1 are both 
around 260 seconds.


Sound quality:

Mostly good for mtest01.  page-cluster is 3.  

Not as good, for mmap01.    I'll give a subjective 4 on a scale of 10.  

Ideally I'd have a nice long mp3, and could say the test took 18:27 and mp3 
played 15:10 (or whatever).  


One other note for people who do similar tests.  mp3blaster skips less
when compiled with glibc-linuxthreads-2.2.4 (default configure) than with 
pth-1.40.  glibc-linuxthreads uses more memory and creates more processes
though.  All of my tests were with a glibc threads mp3blaster.

I'm looking forward to playing with Andrea's new knobs in /proc/sys/vm.

Have fun!
-- 
Randy Hron