No speedup using multiple processors on 10-processor PowerChallenge

Ron Meleshko ron at biochem.ualberta.ca
Thu Nov 16 11:53:35 EST 1995


Kay Diederichs (kay at loop2.biologie.uni-konstanz.de) wrote:
: Dear X-PLOR users,

: for a couple of weeks, I have access to a SGI PowerChallenge (10 R8000
: processors, 1Gb memory, running IRIX 6.0.1). I compiled and linked X-PLOR 3.1f
: after downloading the relevant files from xplor.csb.yale.edu.
: Using the benchmark1 and benchmark2 scripts, I came up with the following,
: surprising results:

:           MP_SET_NUMTHREADS     CPU time        wall-clock-time
: benchmark1  1                 76.1 seconds          95 seconds
:             2                 55.8                 286
:             3                 45.7                 743
:             4                 44.4                 861

: benchmark2  1                 20.4                  27
:             2                 19.2                 306
:             3                 18.1                 280
:             4                 17.6                 277

: I should add that the system load was between 30 and 50 during the tests, but
: the results seem pretty reproducible.
: Does this mean a $$$$ computer does not complete a single job in less time
: than, say, a 250MHz R4400 Indigo2 workstation? Can I do anything about it? 
: ---
: Kay Diederichs                    
: email: kay at loop2.biologie.uni-konstanz.de     Tel: +49 7531 88 4049  Fax: 3183
: Fakultaet fuer Biologie, Universitaet Konstanz 
: Postfach 5560 (M656), D-78434 Konstanz, Germany


Not having access to an MP system, the results may or may not
be surprising.  I took a quick look at the files at xplor.csb.yale.edu
and it seems that all that is there is an MP fft.  This will buy
some speedup, but keep in mind that depending on the benchmark
(we don't keep them online) the amount of actual execution that
will benefit from parallelism can be quite limited.  Simulated
annealing is a time-dependent algorithm and is unlikely to
benefit from multiple processors.  On the other hand, something
like a Monte Carlo algorithm can fall into the "embarassingly
parallel" category.

That said, I can say that I have seen benchmarks on the SGI Power
Challenge (or was it just the challenge) that did exhibit reasonable
speedups for molecular dynamics runs.  However, one of the problems 
with the MP systems is that you often have to "customize" the 
code to achieve good performance.  I do not know if the Xplor source
includes parallelization directives; the compilers can recognize
certain cases where parallelism is possible, but not all (and 
not many :-)).

Regards,

Ron Meleshko
MRC Group in Protein Structure and Function
University of Alberta




More information about the X-plor mailing list