2.1.35 - scheduler bug

Nathan D Spande (nspande@gac.edu)
Thu, 17 Apr 1997 15:40:23 -0500 (CDT)


Hello all. I think I have found a bug in the scheduling code
for 2.1.35. Of course, the situation in which it arises is
a bit odd, so we may not care.

I built 2.1.35 and installed it on my office machine, a P5-133.
I'm testing some of the new features (nfs and autofs in particular)
but in my machine's off-time I run my genetic programming research.
This stuff is a real CPU hog, taking around 99% of the processor
for hours on end if given the chance.

I noticed that when it was running the system behaved like Windows,
not allowing me to do anything else for long periods of time. I
could suspend the process and then things were cool, but while it
was running, I had to wait for minutes before even getting the
"Password: " prompt while logging in.

I suspected that it was a bug in the scheduler, so I rebuilt the
kernel as a non-SMP one, and the problem disappeared. SO...
running an SMP kernel on a uniprocessor seems to kill the scheduler.

One option would be to call this a feature, forcing non-SMP people
to run non-SMP kernels. My opinion, however, is that SMP should
work properly for uniprocessor folks as well. Anyway, this seems
like the kind of thing that could cause problems for SMP folks as
well, and should at least be looked into.

Nathan Spande
nspande@gac.edu