Re: octave + mkl

[Joshua Dillon]

All interesting questions....  Unfortunately I was just playing with
MKL at home--I cannot use it at work due to its restrictive license.
Which is all the more reason why your questions are good and worth
answering.  If I have some extra cycles I'll look into these other
packages.  Frankly the only reason I tried to link octave against MKL
is because `ldd MATLAB` told me that's what they do and I would like
to see octave be as fast as (and as complete) as Matlab.

I am more interested in developing an octave toolbox for machine
learning.  It is my plan to do this over the coming months/years.
This effort would be nicely complemented by parallelized linalg ops.

Cheers,

Josh

On Wed, May 16, 2012 at 9:49 PM, Jordi Gutiérrez Hermoso
<address@hidden> wrote:
> On 16 May 2012 17:55, Joshua Dillon <address@hidden> wrote:
>> Sorry I didn't intend to start a fight.  ...Just thought "It was
>> useful to me and may be useful to others."
>
> Right, what would be far more useful to others in the long run is if
> we could figure out what sort of computatons MKL is performing and
> implementing similar ones in LAPACK, FFTW, OpenBLAS, ATLAS or another
> free BLAS, instead of promoting non-free software and opaque magic.
>
> Can you help us with that?
>
> What is MKL doing? Which BLAS were you comparing with MKL? Is MKL
> parallelising better? Do you see all of your cores being used better
> with MKL? Is it using hardware instructions better on the Intel
> hardware? If you try it on AMD hardware, is MKL also deliberately
> pessimising like icc does, or do you also see an improvement?
>
> Are you really seeing drastically better performance with eigs? All of
> the heavy lifting in eigs is done in ARPACK, and off the top of my
> head, it shouldn't make too much of a difference which BLAS you're
> using, although of course there are BLAS calls in ARPACK.
>
> There is a parallel version that uses OpenMPI, PARPACK. Are you seeing
> much of an improvement with that one? I can't recall right now if
> PARRPACK is a complete drop-in replacement or if we would have to
> modify liboctave/eigs-base.cc in Octave to make use of the parallel
> version. There's also the issue that I'm not sure if OpenMPI can work
> easily on a single multicore box.
>
> I do hope Google hasn't tried to quell the rebellious and playful
> hacker in you and you can still dedicate some of your 20% to these
> interesting questions. :-)
>
> TIA,
> - Jordi G. H.



-- 
Joshua V. Dillon, Ph.D.
Software Engineer, Google Inc.
650-404-7120 / address@hidden / http://www.almostsure.com
"If you work really hard and you're kind, amazing things will happen."
   -- Conan O'Brien