Skip to contentSkip to conversation

peculiar scheduler execution pattern

question Eric Weigle 1 participant 1 message Feb 1, 2002

lore
Share & tools
lore
#0 lore raw

Good morning-

Summary: I am measuring how long the kernel spends in schedule(). On a dual CPU machine, the time for one CPU is lower than the other. Regularly (about every second) the results switch. Why is the time spent in schedule() longer on one CPU than the other and why do the results switch periodically? I vaguely remember seeing something about this on the list recently (with all the O(1) discussion), but searched the archives to no avail.

Details: The machine is a dual 400 MHz Pentium II running Debian testing with a 2.4.17 SMP kernel. I am using an instrumentation package we are developing [1] to save a timestamp at the beginning and end of kernel/sched.c:schedule(). The time source is get_cycles(). Other than an empty infinite loop program, the usual daemons and my ssh shell, nothing is running on the machine. (When not running the CPU bound process, uptime reports a load of 0.00.)

A graph [2] of the time spent in schedule as a function of time shows one CPU spends less time in schedule than the other one (approximately 2 usec and 6 usec respectively). After about a minute they switch places.

Why is the time spent in schedule() longer on one CPU than the other and why do the results switch periodically?

Any information, links, etc. would be appreciated (and I'm on the list, so a CC is not required).
-Eric

[1] http://www.lanl.gov/radiant/website/research/measurement/magnet.html
[2] http://public.lanl.gov/mkg/cpubound_normal_ctx_time.png

Signature · 5 lines
--------------------------------------------
 Eric H. Weigle   CCS-1, RADIANT team
 [email protected]     Los Alamos National Lab
 (505) 665-4937   http://home.lanl.gov/ehw/
--------------------------------------------