On Fri, Jun 24, 2005 at 09:33:20AM -0400, Jesse Harvey wrote:
>I recently ran the HLP Benchmark and was hoping that someone would be
>willing to help me interpret the results, including fidning RMax...
please ask LAM/MPI questions on the LAM list.
however,
http://www.netlib.org/benchmark/hpl/tuning.html
your goal is to tweak things until you get the highest GFLOPS
the results are meaningless unless you crank up the problem size (Ns)
so that you fill at least 1/2 of ram (use top and ps to check) and
preferably 90% of ram - just before the machine start swapping.
your Ns's are ridiculously small.
we find NBs of 80 or 200 work pretty well. Other parameters that have
a big effect are P,Q, and PMAP. Most of the rest don't seem to have a
huge effect.
cheers,
robin
|
