I ran some load tests on the 2.2.0 release kernel on my dual
PentiumII@450 128MB machine. I used a complete clean build of the 2.2.0
release source tree in one, two and three separate processes, using make
-j N where N varied from 1 to 10 (not all values were tried). (Is
anyone interested in the "time" output from some of the trials?)
I watched the system's performance during this excruciating load (load
average nearly 30 for N=10!). It appeared to maintain a reasonable
balance between cache and user data. In fact, the only processes
actually swapped where the ones that were blocked for long periods
(things like atd). This last bit is likely due to the fact that even
this prodigious load was not enough to truly require more than 128MB in
all working sets plus overhead. I will perform other trials that force
actual swapping of computing processes sometime soon.
BTW, one conclusion I can draw from the results of the kernel builds is
that make -j N where N is less than 4 is slower than it has to be - at
least on my configuration (dual PentiumII@450 with lots of RAM and a
fast UltraDMA disk). Much larger than N == 4 is not particularly
useful.
I also tried interaction during this load test. I was running X11R6
(whose RSS is 31 typically MEGABYTES, dammit!). I ran Netscape (also
very HUGE) and browsed a bunch of web sites. All with very little
impact from the huge compute load going on in the "background".
Another test I performed unfortunately fared less well. I was running
with only 128MB of swap. When I substituted a single "make -j" with no
limit on the number of processes to run, I got a bunch of
"resources" messages and the machine wedged. From the outside it was
dead - no network response and nothing I did to the keyboard or mouse
had any effect. I was not able to get more info via the sysrq mechanism
(does this even work when you're running X?). I have not tried this on
a uni-processor configuration nor have I gone further in my exploration
of the problem. That's for later this week, perhaps.
I would very much like particularly to applaud the efforts of Andrea and
Steven and Linus for the VM system tuning. This system seems rock solid
(except for swap running out issue above). My experiences with VM
systems in my career give me a great appreciation for their blood sweat
and tears and the terrific results therefrom. You guys are GODS (and a
GODDESS)!
[waxing philosophical] The work we do on Linux is work to prevent the
Human Race from being stuck with Windows for a LONG TIME as there is no
other credible competition. I worked HARD at Apple for 9 years trying
to prevent this "Windows is the only game in town" scenario from
happening, but it happened anyway. I frequently clench fists in
frustration. Linux is the bright light in the heretofore deepening
gloom. It sucks less every day!
Cheers,
a
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/