P4 slowdown.

[Peter Soendergaard <soender@cs.utk.edu>]

Hi everyone.

I have been benchmarking a Pentium IV chip for some time now, but I have
sometimes gotten very illogical and inconsistant results when I did long
series of tests, and I have just now discovered what went wrong. The
Pentium IV has a safety device that slows the chip down if it gets to hot.
It has been mentioned quite a lot on the net recently, but most sources
concluded that it was unlikely to happen. It does happen however, on the
machine I benchmark, after about 10 minutes of running at full speed. Take
a look a the included .eps file. This is simply a run of 200 identical
tests of my fastest SSE2 double precision code. After about 10 minutes of
running at full speed, the machine start slowing down to about 3/4 the
normal speed.

I have repeated the test several times, and I get the same result every
time. I have tried inserting a 4 seconds pause in between each test, at it
solves the problem, so I see no other logical explanation that the thermal
safety device in the chip.

If anyone wants to repeat my test the can grab my codegenerator at
www.cs.utk.edu/~soender/atlas/src/ps_atlas.tgz

Untar it on top of your ATLAS tree, go to ATLAS/tune/blas/gemm and execute
runhot.py (requires python) and look at the file runhotdata. First column
is seconds, second column is mflops. If you have gnuplot installed execute
'gnuplot gnuplotrunhot' and look at runhot.eps

Cheers,

Peter.

runhot.eps