Howdy,

I tested two systems, both with various kernel patch configurations, with
the methods mentioned in my previous email to check how precise the
scheduling is. 

*** Systems tested ***:

A. Amd Athlon K7 600Mhz with 2.4.18 kernel, rtl-3.1, no X running, no
activity at all (0 users logged in, only cron jobs), XFS filesystem

B. P4 Intel, 1.7GHz 2.4.18, rtl-3.1, variations of X running/not
running, XFS filesystem or normal ext2 fs. 

Tests were done for about 12 hours, or more, so I have almost ~10^5
data point per test.

All my test results are at the following html page:
http://www-cfa.harvard.edu/~gbakos/rtlinux-stat/

See especially the README file for explanations, and the *.png figures
with histogram plots of the "max" column output of the rtl distro
"monitor" program.

*** Conclusions ***:

1. There is a very extended tail to the distributions, with values as
big as 10^5 .. 10^6! This is true for all configurations, i.e. even for
plain ext2fs and rtlinux. See the *.long.png files. 

Q : Why is this? Is the output of monitor bad, or really there is such
    a delay in scheduling? I might misinterpret the output of monitor,
    and this might not be a problem at all.

2. Running X and putting load on the system obviously increases the
bigger max hits, see e.g. rtlstat.1.short.png and rtlstat.1.long.png

3. It is easy to see that the P4 performs better that the K6 (what a
   surprise:). The difference in the median is ~1000 units [nanosec?].
   See rtlstat.2.*.png

4. I definitely see a difference in the long-tail distribution of max
   hits between a system first patched with rtl and then with xfs, as
   compared to a system with opposite order. The xfs-rtl is better. 
   Without xfs (but ext2( there are slightly less high hits, but no
   significant difference. See rtlstat.[36].*.png 


Overall conclusion is that the median values of the distributions are
sth like less than 10000 [unit?] and the best is 2000. Every distrib
has an extended tail with values at least 10^5. 

If I understand better what is going on, I will proceed, and test some
other configurations.
Comments are welcome. And especially answers. 

Cheers
Gaspar


_______________________________________________
Rtl mailing list
[EMAIL PROTECTED]
http://www2.fsmlabs.com/mailman/listinfo.cgi/rtl

Reply via email to