The Need for Speed

Peter Beerli beerli at genetics.washington.edu
Fri Sep 17 11:43:00 EST 1999


Hi all,

Mary Kuhner wrote:
> > The PHYLIP documentation has such benchmarks for some older machines;
> > you could pick up the two data sets (a "best case" data set with
> > very high information content, and a "worst case" noisy one) used
> > for that, and get a bunch of early-machine benchmarks for comparison.
We need to check how those old PHYLIP benchmarks were produced,
for a good comparison among machines one needs to "aggressively" optimize
during the compile (eg: gcc -g file.c vs gcc -O3 -funroll-loop ....
produces up to 3x slower code) so that one can compare speed of machines
for a given compiler.
I think Joe has a newer set of benchmarks in the works.
 
Altivec on G4: If somebody is willing to try on a G4 I can compile 
either dnaml or (preferred) a program from the LAMARC package
(http://evolution.genetics.washington.edu/lamarc.html) using Metrowerks
Altivec support, and so we could see if the metrowerks compiler can
improve the speed without suffering us through many specific code changes.

Julian wrote: 
> What I would like to do is post both data sets to the newsgroup (see below),
> hopefully some people will run them for us, and I can gather a list of
> Machine types
> (is PAUP* 4.0 quicker on IBM or MAC, what about Mainframe UNIX machines),
> 'clock speed' , operating system (e.g. Mac OS 7.5.5 -vs- Mac OS 8.5, etc..),
> and how fast the search was completed.
> We should also do this for other searches in PAUP such a ml, etc. and for
> other
> programmes.
I would like to see additional comparisons, PAUP is certainly a good candidate,
although I would prefer programs that allow me to compile it myself (and optimize to my taste),
what about dnaml?
My prognosis for the length of analysis with paup would be: Dec Alpha 21264/DUNIX < Intel III500Xeon512/LINUX <
Dec Alpha 21164/500MHZ <<< Mac/Win 
Where would be Mac/Altivec, Metrowerks/LINUX ?

Julian wrote: 
> Posting data sets.- I know some newsgroups are not very keen on 'binaries'
> being posted,
put the data sets on your website and tell us where thye are.


Peter
-----------------------------------------------------------
Peter Beerli <beerli at genetics.washington.edu> 
University of Washington, GENETICS, Box 357360, Seattle, 
WA  98195-7360,  USA. Work:(206) 543-8751,
Home:(206) 527-9906, Fax:(206) 543-0754; GMT+0800.
http://evolution.genetics.washington.edu/PBhtmls/beerli.html






More information about the Mol-evol mailing list