[Beowulf] GPU (was Accelerator) for data compressing

Bruno Coutinho coutinho at dcc.ufmg.br
Fri Oct 3 18:33:59 PDT 2008


2008/10/3 Vincent Diepeveen <diep at xs4all.nl>

>
> On Oct 3, 2008, at 5:45 PM, Joe Landman wrote:
>
>>
>> GPU as a compression engine?  Interesting ...
>>
>> Joe
>>
>>
> For great compression, it's rather hard to get that to work.
> With a lot of RAM some clever guys manage.
>
> GPU has a lot of stream processors, yet little RAM a stream processor.


A tesla have 4GB ram for 240 stream processors. That gives ~16MB for
processor.
But SMID arrays (stream multiptocessors in CUDA) need a lot of threads to
tolerate memory latencies and not get idle.
Something of the order of 512 threads per multiprocessor (tesla have 30), so
we have 4GB divided among 15360, that gives 273k per thread.


>
> Additionally they all have to execute the same code at a bunch of SP's at
> the same time.
>
> So there is a big need for some real clever new algorithm there,
> as the lack of RAM is gonna hurt really bigtime.


You need to get data from machine main memory, compress and send results
back several times.
The bandwidth of pci express today is 8GB/s, so this is the maximum data
rate a gpu can compress.
You can use some tricks like computation and i/o (to main memory)
parallelization, but will be constrained to 8GB/s anyway.


>
>
> Would be a mighty interesting study to get something to work there. It
> allows real
> complicated mathematical functions for PPM functionality.
>
> What's in that GPU soon will be in a CPU anyway, so it benefits the entire
> planet a thing like that.


This will be interesting.
Removing pci express from the path, the gpu will become really a parallel
coprocessor.

With the pressure for a more flexible programming model caused by Larabee,
the coprocessor could become as programable as the main processor and we
will have a processor with few big cores for serial workloads and many cores
for parallel workloads.


>
>
> Where can i ask for funding?
>
> Vincent
>
>
>
>  Cheers
>>> Carsten
>>>
>>
>> --
>> Joseph Landman, Ph.D
>> Founder and CEO
>> Scalable Informatics LLC,
>> email: landman at scalableinformatics.com
>> web  : http://www.scalableinformatics.com
>>       http://jackrabbit.scalableinformatics.com
>> phone: +1 734 786 8423 x121
>> fax  : +1 866 888 3112
>> cell : +1 734 612 4615
>>
>>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20081003/9eee6422/attachment.html


More information about the Beowulf mailing list