Re: Linux Multi-CPU performance


"Scott Hazelhurst" <scott@xxxxxxxxxxxxx> wrote in message
news:lvirs1c31s.fsf@xxxxxxxxxxxxxxxxxxxxxxxx


Dear all,

We would be grateful for any suggestions on a problem....

We've developed an SMP version of a program we had previously
successfully parallelised on a cluster. The results were very
disappointing with no obvious reasons for such poor performance.

After some experimentation we have started to wonder if it's not on OS
bug. I wrote small programs that use almost no memory, no I/O and are
completely compute bound.

Experiment 1: On a single CPU machine
-- base case: time how long the program takes to run
-- run concurrently 4 copies of the
same program. Result: overall wall-time 4 times longer, each process
taking the same time as the base case
Expected

Experiment 2: On a 4-CPU machine
-- base case: time how long the program takes to run
-- run concurrently 4 copies of the same program.
Result: overall wall-time about 60% longer. Each copy of the
program, taking 60% longer than it did previously.


We are running an oldish version of the kernel (not in my control
entirely)
Linux juggernaut 2.6.11-1.1369_FC4smp #1 SMP Thu Jun 2 23:08:39 EDT
2005 i686 i686 i386 GNU/Linux

Is it worth pushing to upgrade the kernel?

Any other suggestions


FWIW, /proc/cpuinfo looks OK (shown below).


------ 8< snip 8<----------------------

Are you sure the 4 concurrent programs are each running on a separate
processor? Maybe you should try setting the CPU affinity for each instance
of the program.

Take a look at this article:
http://www.linuxjournal.com/article/6799



.