BLAS lib performance (was Re: MPI (was Re: BIG libraries ....))

[Fredrik Lingvall]

> > > > Hi, did you happen to build octave linking to it, and if so, what 
> > > > were the effects on performance? Cheers, M.
> >
> > > If you carry this out, please report back on your results (including
> > > performance).
> >
> >
> >
> > I did this some time ago:
> >
> > http://www.octave.org/mailing-lists/help-octave/2004/168
I have done a quick performance test with ATLAS vs. GOTO BLAS on a 
Pentium M machine (DELL D810) and
a dual Pentium 4 Prescott machine (DELL Precision 670) using the code 
below (octave 2.1.71):

T = [];
N = [];

%for n=[500:500:5000 1000]
for n=[500:500:5000]
 A = randn(n,n);
 tic;
   B=A*A;
 t = toc;

 fprintf('Elapsed time (n=%d) %f\n',n,t);

 T = [T t];
 N= [N n];
end

Results:

**************** 2.0 GHz Pentium M 760 (2MB cache), Dell Latitude D810 
laptop, 2 GB ram:

* ATLAS (libatlas_Linux_P4SSE2.so)

Elapsed time (n=500) 0.192438
Elapsed time (n=1000) 1.481611
Elapsed time (n=1500) 4.929739
Elapsed time (n=2000) 11.702435
Elapsed time (n=2500) 22.817465
Elapsed time (n=3000) 39.251930
Elapsed time (n=3500) 62.606858
Elapsed time (n=4000) 93.108703
Elapsed time (n=4500) 132.630548
Elapsed time (n=5000) 181.885659

* GOTO BLAS (libgoto_northwood32p-r1.00.so)

Elapsed time (n=500) 0.165478
Elapsed time (n=1000) 1.260498
Elapsed time (n=1500) 4.224197
Elapsed time (n=2000) 9.870070
Elapsed time (n=2500) 19.341520
Elapsed time (n=3000) 33.227217
Elapsed time (n=3500) 52.873640
Elapsed time (n=4000) 78.452588
Elapsed time (n=4500) 112.102844
Elapsed time (n=5000) 152.980586


**************** Dual 2.8 GHz XEON (1MB cache), Dell Precision 670, 3 GB 
ram:

* Treaded ATLAS (libATLAS_Linux_P4ESSE3_2.so)

Elapsed time (n=500) 0.069242
Elapsed time (n=1000) 0.259086
Elapsed time (n=1500) 0.816377
Elapsed time (n=2000) 1.887494
Elapsed time (n=2500) 3.930341
Elapsed time (n=3000) 6.259690
Elapsed time (n=3500) 9.992612
Elapsed time (n=4000) 14.733824
Elapsed time (n=4500) 21.107499
Elapsed time (n=5000) 29.036249
Elapsed time (n=10000) 227.612336

* GOTO (libgoto_prescott32p-r1.00.so)

Elapsed time (n=500) 0.067996
Elapsed time (n=1000) 0.252867
Elapsed time (n=1500) 0.893660
Elapsed time (n=2000) 1.821469
Elapsed time (n=2500) 3.519627
Elapsed time (n=3000) 6.064184
Elapsed time (n=3500) 9.580817
Elapsed time (n=4000) 14.161038
Elapsed time (n=4500) 20.221677
Elapsed time (n=5000) 27.726569
Elapsed time (n=10000) 218.574903

* GOTO (libgoto_prescott32p-r1.00.so) 
with export OMP_NUM_THREADS=1

Elapsed time (n=500) 0.081871
Elapsed time (n=1000) 0.446672
Elapsed time (n=1500) 1.482445
Elapsed time (n=2000) 3.452727
Elapsed time (n=2500) 6.744513
Elapsed time (n=3000) 11.551446
Elapsed time (n=3500) 18.490894
Elapsed time (n=4000) 27.245389
Elapsed time (n=4500) 39.056861
Elapsed time (n=5000) 53.391296
Elapsed time (n=10000) 422.807981

On the laptop GOTO blas was roughly 18% faster than ATLAS and GOTO blas 
was about 4%
faster on the dual XEON. The speedup by using two threads instead one was
roughly 1.9 times on the dual XEON machine (with GOTO BLAS).

Fredrik

BTW. I noticed that matlab utilizes that B=A*A'  (and A'*A) is symmetric 
(A*A' only takes
half the time to compute compared to A*A). Is this functionality 
available in octave as well?





-------------------------------------------------------------
Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:  http://www.octave.org
How to fund new projects:  http://www.octave.org/funding.html
Subscription information:  http://www.octave.org/archive.html
-------------------------------------------------------------