[Wien] Speed versus size (e.g. RKMAX)

[Torsten Andersen]

Just guessing... is the LINPACK benchmark showing something similar?

I assume that you have more than 1GB of physical memory.

It could be the CPU design or cache optimization. I do not know the 
compiler, but the Sun compilers (for sparc) gives much better speed if 
the cache size and type is specified.

It is also "well known" that the newer "prescott" P4 core needs at least 
twice the cache size to be within 5% of the speed of the older 
"northwood" P4 core, so assuming it is a design issue seems reasonable - 
maybe the cache register runs out and the number of "cache misses" increase.

Does it change if you recompile under a compiler that supports 
"prescott" P4s explicitly?

All this is, of course, speculation, but maybe it can give a hint or two.

Best regards,
Torsten Andersen.


L. D. Marks wrote:
> We've run across something odd which someone might know about, and the 
> answer may well be generally useful for Wien2k calculations. With our 
> older cluster (3.4 P4 with 512L2) once the total size of the calculation 
> (i.e. top) gets above about 1G the speed (mflops) drops. We have a newer 
> cluster (3.4 P4 with 1024L2) which is about 5% slower for the benchmark, 
> but shows a similar drop in speed now for a much smaller size (about 
> 500Mb). I suspect that this has to be some sort of "tuning" parameter in 
> the kernel, but don't know which one and which way to start adjusting. 
> Any suggestions?
> 
> Note: if you have an old email address for me, please note that "nwu" has
> been changed to "northwestern".
> -----------------------------------------------
> Laurence Marks
> Department of Materials Science and Engineering
> MSE Rm 2036 Cook Hall
> 2220 N Campus Drive
> Northwestern University
> Evanston, IL 60201, USA
> Tel: (847) 491-3996 Fax: (847) 491-7820
> email: L-marks at northwestern dot edu
> http://www.numis.northwestern.edu
> -----------------------------------------------
> 
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> 

-- 
Dr. Torsten Andersen        TA-web: http://deep.at/myspace/
AG Hübner, Department of Physics, Kaiserslautern University
http://cmt.physik.uni-kl.de    http://www.physik.uni-kl.de/