Re: [Fsuk-manchester] parallel systems (was: Open ARM GPU drivers)

[Bob Ham]

On Tue, 05 Feb 2013 17:53:07 +0000, Dave Love <address@hidden>
wrote:
> Bob Ham <address@hidden> writes:
> 
>> Also interesting, and could be useful.  I find it noteworthy that the
>> library targets multi-core CPUs, rather than GPGPUs.  I've viewed
OpenMP
>> as
>> the de facto method for parallelism on CPUs and OpenCL for GPGPUs but
on
>> reflection, OpenMP seems only like a quick and easy way to get instant
>> parallelism.
> 
> If only!
> 
> Most HPC is actually done with MPI

Yes, I should have qualified my statement really; I was referring to
contexts outside of HPC.  I'm aware that MPI rules the roost within the HPC
world.

For the right kinds of problems, I think modern GPUs really do provide a
supercomputer-on-a-chip and I'm continually dumbfounded by their leaps in
performance.  They allow amateurs or researchers without big budgets, to do
the kind computations that were previously reserved for the realm of HPC,
using commodity hardware on their desktop.  It's a brave new world :-)

I needed parallel computing to calculate a particular statistic for the
analysis of data collected for my (non-computing) MSc thesis. 
Unfortunately, the algorithm for the calculation had a high order of time
complexity.  Having implemented the algorithm in OpenMP for my dual-core
Core2 processor, I was dismayed to find that it would take a few months for
the computation to complete, which was impractical with a deadline looming.
By comparison, my experiments with OpenCL on a GPU hinted that it should
take only a number of days.  Unfortunately, due to said deadline I didn't
have time to implement it (and so wasn't able to use the statistic in my
analysis).

This experience really showed me the extraordinary power that GPU
computing puts in the hands of those without access to HPC.  It's these
kinds of users that I was focussing on above.


> Basically you don't want to work at the CUDA/OpenCL, or even OpenMP,
> level if you can reasonably avoid it, and the interesting systems are
> networked at multiple levels.  One free software system addressing that
> is <http://runtime.bordeaux.inria.fr/StarPU/>.

Fascinating.  I may look into that.


> [Why don't we have an INRIA?]

I'll take a stab at answering :-)  The French have quite a strong sense of
communal society; "liberté, égalité, fraternité", etc.  By comparison, the
English seem to have quite a strong sense of greed, something the
government considers a good thing.  So, instead of INRIA, we have
for-profit labs like Microsoft Research Cambridge:

  http://research.microsoft.com/en-us/labs/cambridge/default.aspx

and IBM Hursley:

  http://www.linkedin.com/company/239147?trk=saber_s000001e_1000

Alas.


-- 
Bob Ham <address@hidden>

for (;;) { ++pancakes; }