|
"David " <djones.nospam@remove.mcmaster.ca> wrote in message
news:feqvbe$s88$1@fred.mathworks.com...
> "Penny Anderson" <penny@mathworks.com> wrote in message <faf9ua$ljk
> $1@fred.mathworks.com>...
>>
>> You probably aren't using the math functions in MATLAB that are
>> multithreaded. These include the BLAS-based linear algebra routines
>> including matrix multiplication, and the element-wise operators.
>>
>> Penny Anderson
>> The MathWorks, Inc.
>
>
> The new R2007b multithreading doesn't seem to lead to any improvement on
> a a DUAL 2.5 GHz G5 Mac (PowerPC) when running the MATLAB benchmak
> ("bench"). Clearly that benchmark must be designed to test the BLAS
> routines
> and element-wise operations, ... or else it is a useless benchmark. There
> was
> very little (if any) difference when switching between multi-threading on
> or
> off.
>
> What gives?
>
> What example or benchmark can a MATLAB user run to verify whether there is
> any multi-threading happening, and measure the benefit (for instance, on a
> quad-core CPU, or multiple CPUs).
>
> -- David Jones
David,
bench is an old benchmark from pre-multithreading days. In fact, only the LU
benchmark is a really good test of the new multithreading features, in
particular the BLAS routines underlying the LAPACK routine DGETRF called
within LU. On Mac/PowerPC, multithreaded BLAS is ALWAYS on so even though
you think you are toggling between single-threaded and multithreaded
performance, in fact you are always seeing multithreaded BLAS kicking in. As
an aside, the SPARSE benchmark ends up calling BLAS routines as well, and
you are always seeing those in multithreaded variants on PowerPC as well.
Try to run the multithreaded computation demo which will also run
element-wise operations with and without multithreading. Those should show
performance differences from single-threaded to multi-threaded runs.
http://www.mathworks.com/products/matlab/demos.html?file=/products/demos/shipping/matlab/multithreadedcomputations.html
Penny.
|
Feed for this Thread