decreasing #define HZ in the linux kernel for CPU/memory bound apps ?
hahn at physics.mcmaster.ca
Tue Apr 16 13:35:54 EDT 2002
> Would it be interesting to decrease the #define HZ in the linux kernel
> for CPU/Memory bound computationnal farms ?
I'm guessing you're unaware that compute-bound processes
actually get multiple 10ms slices (200ms or so, as I recall,
but I'm remembering a discussion from 2.3.x days. Ingo's new
scheduler probably preserves this limit.)
> I mean we very often have only one running process eating 99% of
> the CPU, but we (in fact I) don't know if we loose time doing context
> switches ....
think of the numbers a bit: it's basically impossible to buy
a <1 GHz processor today, so you're getting at O(100M) instrs/HZ.
if you're cache-friendly, you'll probably have >1 instr/cycle,
so scale the number appropriately. perhaps you're worried about
cache pollution? the kernel's footprint is fairly small, probably
<4K or so for timer-irq-scheduler-nopreempt. since a null syscall
is ~1 us or ~1000 instrs, and the work is about the same, I really
don't think there's anything to worry about.
there are people who run HZ=1024 or higher on ia32; I don't personally
think they know what the heck they're doing, but they like it, and
don't report any serious problems.
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf