Thread Subject: Multi-core and distributed computing toolbox?

I would like to get some advice about multiple core/
processor computers and using the Distributed Computing
Toolbox. For example, I am considering purchasing a Mac Pro
with two quad-core processors. Would the DC engine be able
to utilize all 8 cores (8X speedup), or would it be limited
to using a single core on each processor (2X speedup) - or
perhaps, worst case, just a single core on one of the
processors (no speedup)? It's not clear to me how this
works. Thanks!

BTW, the speedup increases in () are assuming the bulk of
the program can be made "embarassingly parallel".

"Evan Zamir" <zamir.evan.nospam@gmail.com> schrieb im Newsbeitrag
news:fphhqp$fnc$1@fred.mathworks.com...
>I would like to get some advice about multiple core/
> processor computers and using the Distributed Computing
> Toolbox. For example, I am considering purchasing a Mac Pro
> with two quad-core processors. Would the DC engine be able
> to utilize all 8 cores (8X speedup), or would it be limited
> to using a single core on each processor (2X speedup) - or
> perhaps, worst case, just a single core on one of the
> processors (no speedup)? It's not clear to me how this
> works. Thanks!
>
> BTW, the speedup increases in () are assuming the bulk of
> the program can be made "embarassingly parallel".

Hi,
the Distributed Computing Toolbox is able to use multiple
cores in the following sense: it supports up to 4 local workers.
Each worker is MATLAB instance running and they don't
care if they are on physically different machines or using
different cores on the same machine. So the speedup would
be 4 (or of course slightly below).
Using the Distributed Computing Engine the limitation of 4
is dropped, so using the Distributed Computing Toolbox
together with the engine you can use all of your eight cores.

Titus

Hello,

Also, you can try using a combination of explicit multiprocessing (DCT
workers) and implicit multiprocessing (threads in MATLAB).

For example, you could use 4 MATLAB workers, each of which calls
maxNumCompThreads(2). That might allow you to utilize all 8 cores on your
desktop.

For more information see
http://www.mathworks.com/access/helpdesk/help/techdoc/matlab_prog/brdo29n-1.html#brdo29n-2

Arjav


"Titus" <titus.edelhofer@mathworks.de> wrote in message
news:fphk0s$jik$1@fred.mathworks.com...
>
> "Evan Zamir" <zamir.evan.nospam@gmail.com> schrieb im Newsbeitrag
> news:fphhqp$fnc$1@fred.mathworks.com...
>>I would like to get some advice about multiple core/
>> processor computers and using the Distributed Computing
>> Toolbox. For example, I am considering purchasing a Mac Pro
>> with two quad-core processors. Would the DC engine be able
>> to utilize all 8 cores (8X speedup), or would it be limited
>> to using a single core on each processor (2X speedup) - or
>> perhaps, worst case, just a single core on one of the
>> processors (no speedup)? It's not clear to me how this
>> works. Thanks!
>>
>> BTW, the speedup increases in () are assuming the bulk of
>> the program can be made "embarassingly parallel".
>
> Hi,
> the Distributed Computing Toolbox is able to use multiple
> cores in the following sense: it supports up to 4 local workers.
> Each worker is MATLAB instance running and they don't
> care if they are on physically different machines or using
> different cores on the same machine. So the speedup would
> be 4 (or of course slightly below).
> Using the Distributed Computing Engine the limitation of 4
> is dropped, so using the Distributed Computing Toolbox
> together with the engine you can use all of your eight cores.
>
> Titus
>

Thanks, for your replies. That very much helps.

Hello Everybody,

In principle I have the same intention than Evan Zamir. I am
trying to use a server with 8 processors to carry out
distributed computation. Apparently the programs work well
when I use them in a normal desktop computer.

However, I realize that the total computing time is almost
the same than computing with one processors of the same
server. Moreover, the functions createTask(...),
set(Pop,'FileDependencies',{'Dist.m'}), waitForState(...),
and getAllOutputArgument(...) take 30%, 15%, 11%, and 9% of
the total computing time respectively!!!

Does anybody know any reason for this poor performance with
distributed computation in my case?

I thank you in advance any suggestion or advice.

Regards,

Diego Midence