 |
 | Home | Contact Us | FAQ | Search & Site Map | Link to Us |
 | Sign In | Join | Other 45 Sites in Network |
Math Forum / Math Software / MATLAB / July 2008parfor loop running same speed as for loop on Mac
I'm using a Mac with 2, 3 GHz quad-core intel xeon processors, running Mac OS-X. When I run the following short code: matlabpool local 4 tic; parfor i=1:2000 rand(i); end toc; matlabpool close tic; for i=1:2000 rand(i); end toc; matlabpool close the parfor loop clocks at ~73 seconds while the for loop clocks at 72 seconds. If i change the value of i much greater or smaller, the effect is similar, the parfor loop runs 1-2 secs longer (I assume because of the time it takes to distribute the for loop). My first question is how can I make parfor function much faster (as it should) than a regular 4. second, If i try running matlabpool local 8, I get the error message: ??? Failed to create and submit a parallel job. This is caused by: Error using ==> distcomp.localscheduler.pSubmitParallelJob at 17 A job submitted to a local scheduler can only use up to 4 labs is there a way to get matlab to use the full power of this computer? > I'm using a Mac with 2, 3 GHz quad-core intel xeon processors, running Mac > OS-X. When I run the following short code: [quoted text clipped - 21 lines] > similar, the parfor loop runs 1-2 secs longer (I assume because of the time it > takes to distribute the for loop). Strange, I see similar behaviour on one of our Macs, but not on a 64-bit Linux machine. The problem that you're executing there looks like it's very memory-bandwidth intensive (you're allocating, filling, and then de-allocating about 16 GB in that loop), so perhaps you are completely memory-bound. PARFOR works best on problems that are CPU-bound, like: >> tic, for ii=1:100, x(ii) = max( eig( abs( rand( 400 ) ) ) ); end, toc Elapsed time is 38.286138 seconds. >> tic, parfor ii=1:100, x(ii) = max( eig( abs( rand( 400 ) ) ) ); end, toc Elapsed time is 9.935093 seconds. > My first question is how can I make parfor function much faster (as it should) > than a regular 4. [quoted text clipped - 7 lines] > > is there a way to get matlab to use the full power of this computer? Currently, the local scheduler is limited to 4 MATLAB workers. To use 8 workers would require a MATLAB Distributed Computing Server licence: http://www.mathworks.com/products/distriben/ Cheers, Edric. Edric M Ellis <eellis@mathworks.com> wrote in message <ytwzlo15b1v.fsf@uk-eellis-deb4-64.mathworks.co.uk>... > > I'm using a Mac with 2, 3 GHz quad-core intel xeon processors, running Mac > > OS-X. When I run the following short code: [quoted text clipped - 53 lines] > > Edric. Edric, Thanks for the tip on the Distributed Compouting Server. For my initial problem, what you said makes since, except on my duel core PC the code I'm running runs faster using parfor, just not on the Mac. This makes me think something might be wrong with the configuration on the mac. As an example, on my PC if i did parfor i=1:100 fid = fopen(['test_' int2str(i) '.bin'], 'w'); fwrite(fid, rand(i)); fclose(fid); end with 4 workers, files would appear in the directory 4 at a time (because the parfor is being split up 4 ways), however if I run this same code on the mac, with 4 workers, files show up 1 at a time, at the same speed as a for loop. -mike > Edric M Ellis <eellis@mathworks.com> wrote in message > <ytwzlo15b1v.fsf@uk-eellis-deb4-64.mathworks.co.uk>... [quoted text clipped - 93 lines] > > -mike Hi Mike, does your call to matlabpool indeed starts the MATLAB workers? I.e., looking for the running processes, do you indeed see 4 additional MATLABs being started with matlabpool? Titus > [...] This makes me think something might be wrong with the configuration on > the mac. As an example, on my PC if i did [quoted text clipped - 9 lines] > if I run this same code on the mac, with 4 workers, files > show up 1 at a time, at the same speed as a for loop. Unfortunately in R2008a there isn't really a nice way of checking the health of the matlabpool. Here's a few slightly unpleasant things that might shed some light: pctRunOnAll system_dependent( 'getpid' ) and !ps -c | grep MATLAB They should show several different PIDs, which you should be able to check with what's actually running on your system. If the "pctRunOnAll" looks OK (i.e. returns several different PIDs, and matches the "ps" thing), then you could also try: parfor ii=1:100 x(ii) = system_dependent( 'getpid' ); end unique( x ) (There will be less hacky ways of doing this stuff in the next release of MATLAB/PCT). If that doesn't look right - i.e. there's only one PID (probably the PID of the client MATLAB), then there is a problem. Your best bet may be to contact The MathWorks support to help you from there. Cheers, Edric. Edric M Ellis <eellis@mathworks.com> wrote in message <ytwr69b6ca8.fsf@uk-eellis-deb4-64.mathworks.co.uk>... > > [...] This makes me think something might be wrong with the configuration on > > the mac. As an example, on my PC if i did [quoted text clipped - 37 lines] > client MATLAB), then there is a problem. Your best bet may be to contact The > MathWorks support to help you from there. Edric, Your comment about 4 workers being allowed without a distributed liicence interests me. I could not find any concrete description of this (probably version-dependent). Can you point me? - Steve > Edric M Ellis <eellis@mathworks.com> wrote in message > <ytwr69b6ca8.fsf@uk-eellis-deb4-64.mathworks.co.uk>... [quoted text clipped - 61 lines] > > - Steve Hi Steve, the Parallel Computing Toolbox allows for up to 4 local workers running on the same machine, see http://www.mathworks.com/products/parallel-computing/description1.html and there the image at the bottom with the italics text to the right. Titus "Titus" <titus.edelhofer@mathworks.de> wrote in message <g6pkce$56d$1@fred.mathworks.com>... > > Edric M Ellis <eellis@mathworks.com> wrote in message > > <ytwr69b6ca8.fsf@uk-eellis-deb4-64.mathworks.co.uk>... [quoted text clipped - 68 lines] > http://www.mathworks.com/products/parallel-computing/description1.html > and there the image at the bottom with the italics text to the right. I see. I didn't realise you needed to pay for the parallel computing toolbox and then more for the distributed computing thing. I kind of assumed they were one and the same. |
Reply to this Message |
 Sign In
 Join
 My Latest Posts My Monitored Threads My Blog My Photo Gallery My Profile My Homepage Start New Thread Enable EMail Alerts Rate this Thread
|
©2010 Advenet LLC Privacy Policy - Terms of Use
This website includes both content owned or controlled by Advenet as well as content owned or controlled by third parties.