From: Axel Farr <Axel.Farr@physik.tu-darmstadt.de>
Newsgroups: comp.parallel.pvm
Subject: SMP-Linux nearly locks up after some hours of operation
Date: Fri, 30 Apr 1999 10:40:37 +0200
Organization: TU Darmstadt
Message-Id: <37296C85.42E50E22@physik.tu-darmstadt.de>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Xref: ukc comp.parallel.pvm:8331


Hello,

I have ported a program making PIC-Calculations to a
master-worker-application under PVM on Suse Linux 6.0. The master spawns
four worker tasks on a SMP-PPro-Machine with 2x400MHz and some 256
MBytes of RAM. Nearly all of the RAM is used for the calculation.

After an hour of operation, "top" and other utilities display a load
average of more than 6 instead of 4 when the calculation got started,
and the cpu load is around 100% system and 100% user istead of nearly
200% user when the calculation is started.

If I look at the process list, the calculation tasks show status "S" or
"D" instead of "R" when the calculation is started.

The most obvious change is, that the system behaves very lethargic in
that state. When the calculation is started, the system is much more
agile. Utilities with only few computations like "top", "xeyes" "xclock"
etc. all show a cpu usage of some 20%.

I am using PVM 3.4beta6 and the behaviour occurs with kernels 2.2.1 and
2.2.5.
The behaviour seems independend from the runlevel, it occurs with and
without runnung the XFree86 (although when running, X seems to eat up
100% of CPU time).

Axel Farr