Multicore - Parallel processing on multiple cores

This package provides some functions realizing parallel processing on multiple cores/machines.

Description

The latest Matlab releases (starting with R2007a) include support for multiple cores. However, Matlab will only take advantage of multiple cores in evaluating certain computations like an FFT or a FIR filtering operation. Matlab will never be able to determine if, for example, consecutive function calls in a for-loop are independent of each other.  
 
With this package, I provide some MATLAB-functions realizing parallel processing on multiple cores on a single machine or on multiple machines that have access to a common directory.  
 
If you have multiple function calls that are independent of each other, and you can reformulate your code as  
 
resultCell = cell(size(parameterCell));  
for k=1:numel(parameterCell)  
  resultCell{k} = myfun(parameterCell{k});  
end  
 
then, replacing the loop by  
 
resultCell = startmulticoremaster(@myfun, parameterCell);  
 
allows you to evaluate your loop in parallel in several processes. All you need to do in the other Matlab processes is to run  
 
startmulticoreslave;  
 
No special toolboxes are used, no compilation of mex-files is necessary, everything is programmed in plain and platform-independent Matlab. If one of your slave processes dies - don't care, the master process will go on working on the given task.  
 
Please consider that the communication between the processes, which is done by saving/loading temporary files, causes some overhead. Thus, if your function calls only need fractions of a second, the overhead will eat the advantage of the parallelization or can even lead to an increase of computation time. However, the multicore package can do several function evaluations after each other in every process before communicating again. In this way the overhead percentage can be lowered. See the help of function startmulticoremaster.m for more details.  
 
Note: The Matlab multithreading capability (R2007a and higher) might terminate all the advantage gained by using the multicore package. So make sure that you UNcheck "Enable multihreaded computation" under File/Preferences/General/Multithreading in all involved Matlab sessions.  
 
Dicuss with others users here: http://groups.yahoo.com/group/multicore_for_matlab  
 
Keywords: Parallel processing, distributed computing, multiple core.

• MATLAB release: MATLAB 7.5 (R2007b)
• HTML Files: Multicore

Other Files

chompsep.m,
concatpath.m,
datenum2.m,
deletewithsemaphores.m,
displayerrorstruct.m,
existfile.c,
existfile.m,
findfiles.m,
generateemptyfile.m,
getfiledate.m,
gethostname.m,
getusername.m,
license.txt,
mbdelete.m,
mbload.m,
mbtime.m,
multicoredemo.m,
postprocessdemo.m,
regexptokens.m,
removefilesemaphore.m,
setfilesemaphore.m,
startmulticoremaster.m,
startmulticoreslave.m,
tempdir2.m,
testfun.m,
textwrap2.m,
translatedatestr.m,
userfeedback.m

[Michal Kvasnicka]

The demo should be more selfdescriptive. Comparison with single core demo running is very important for parallelization impact evaluation.

[lucia Come]

long due as standard matlab capability  
does it work also with scrits or only with functions ?

[Zhijun Wang]

Parallel processing on multiple cores in a single machine is very useful!

[Zhijun Wang]

Corrections:  
 
I tested the programn to see whether this program can improve parallel processing in a single machine with the following code:  
 
______________________  
N=20;  
for m = 1:10  
tic  
  for z = 1:N  
      testfun(z)  
  end  
 toc  
end  
__________________________  
 
Results show the program do not improve anything comparing with my code in a single machine!  
 
So I rate again

[William Renaud]

Would it be possible to make this compatible with Octave? Due to licensing restrictions it is difficult for many people to have numerous Matlab instances available.

[Michal Kvasnicka]

I am not able to reproduce any improvment from sequential to parallel (multicore) realization of the testing demo code.  
 
I would go as far as to say, that parallel (multi-core) realization is slower then sequential in some cases.

[Michal Kvasnicka]

Wow!!! After the few hours of hard experimentation with MULTICORE I finally learned how to use this code for parallel running on my four core PC. This is realy good and useful tool for distributed computing!!!  
 
1. Informational text in the demo must be extended to detailed description how to run "tesfun" in parallel regime.  
2. Some work must be done to minimize interprocess communication overhead, which may be very itensive (25% of the overall load) in some cases.  
 
Good work!!!

[schena gianni]

on dual-processor Intel Xeon based machines it halfs calculation time i.e. it cuts by a factor 2 !

[Darren 3M]

Brilliant use of the filesystem to share the load. It's not quite 2x increase but on a quick cluster test I got an increase of 4x with 5CPU's so 80% efficient.

[Markus Buehren]

Yes and no: The slave process should create the directory if it is not existing (I have updated that). However, you can start the slave processes whenever you like! You can also interrupt and restart them while the master is running.

[happy matlabuser]

Ran it with 8 processors across 6 machines, and it works beautifully. Unfortunately, if you kill one or more processors, the master processor MUST do that job. Since the jobs are long for my problem, it's better to kill everything in that situation and finish everything at the command line. The master starts at the top of the list, and the slaves start at the bottom. When they meet somewhere in the middle, the master will always redo one of the jobs. I don't think the code will be efficient for a very large number of very small jobs (correct me if I'm wrong); so, I recommend making jobs medium length, (total run time)/(10*(# of processors)) for example. For my problem, these are fairly minor concerns. The author did a great job with this code -- easily saving me hours of work -- thank you!

[Kevin Thom]

Awesome program ... I have run it successfully using anywhere from 4 processors on one machine to 10 processors on 5 machines ... it makes a whole range of computationally expensive projects feasible in MATLAB.

[Huy Bui]

Multicore works quite well overall. I got a bug though. I make a mistake in parameterCell. The slave processes all die because of that. When I tried to exit all slave processes & start again, the same error causing slave processes dead appears again & again.

[David Brown]

Works excellently. It would be nice to see this turned into a fully-guided setup to use this.

[Andrea Soldan]

 
 
very useful and it works excellently.  
much more than the distribuited computing toolbox provided by Matlab (which is very hard to use).  
my SO is Linux, and i'm working with 4 workers ( 2 dual-core processors)  
 

[A. S.]

Thanks, very useful. Synchronization is not optimal (for example, the master shouldn't start working on a task if a slave is already working on it), but still a great program.

[igor scardanzan]

great , just some difficulties to kill the slave : the process persists and CNTR C does not work . one should await the execution end

[Igor S]

An option for avoiding the use of the master core is desirable  
Indeed if one slave process terminates or crushes the entiere process continues but if by chance the iteration that crushes is on the master core all the process is compromised

[Markus Buehren]

> An option for avoiding the use of the  
> master core is desirable  
 
The option is already existing: You can use the input parameter MAXMASTEREVALUATIONS and set it to zero.  
 

[Jun Kim]

This process works very well. For my model estimation, I was able to cut the execution time in a drastic manner. Also Markus was kind and was very responsive to my question.

[uju jbl]

 

[Robert Turner]

Brilliant library. Works like a charm

[M H]

Its absolutely excellent. Am using it now on my quad core machine and am probably going to buy another quad core just to see my models run so quickly. :)

[Bruno Cordani]

Great!!!

[Janos Rohaly]

It seems there is the possibility for master and slave to simultaneously evaluate the same set of parameters. For example, if slave starts evaluating a slow process, master can catch up, and there is nothing to prevent it to start the very same computation since slave hasn't generated the result file yet. There is also a bug in setfilesemaphore.m. dirStruct(k).datenum in line 78 should be datenum(dirStruct(k).date).

[Marcio Sales]

I tested the functions on two dual core machines. I had great gain when paralellizing processing between the processors of dual core machine or between the two machines using a single processor in each. However, I had no significant gain when trying to using both processors on both machines. Is that because the gain of using more processors is being reduced by increased load of data recording when you add more processors/machines? Also, there are times when I get an error in which one of the workers can't delete the temporary files and this seems to happen more frequently when you increase the number of workers. Is anyone having the same issues? My two machines run Vista 64bits.

[Arturo Serrano]

I got the same problem as Rohaly. When master ends the computation, and there is a slave still working, the master computes the remaining job again, yielding an extra iteration.  
The solution is to set MAXMASTEREVALUATIONS, with all its drawbacks, since i understand that this isn't a bug rather than a problem not knowing if the slave got interrupted. 
 
BTW, it works like a charm.

[Arturo Serrano]

 

[Vasilis Kapetanidis]

Thank you! This works just fine and now I have 100% CPU usage on all 4 cores! By measuring the elapsed time it seems that it runs about 3.4 times faster than with a single matlab doing all the work, so that's about 85% efficient on my quad-core 
 
now, if only this could run on a single multi-core machine with only one matlab instance running

[Jordi Arnabat]

Thanks for this great contribution, it's very useful. 
 
Correction: 
I've used in under a grid of computers running different OS: GNU/Linux, Mac and Windows (XP). When the shared folder is on a network computer (not mapped to a local drive, for example: \\servername\sharedfolder); Windows systems fail trying to delete the semaphores, causing the master process run forever. 
 
The solution I found is to slightly modify compsep.m and concatpath.m as follows: 
 
_______________________________________________________ 
function str = chompsep(str) 
unix_sep = '/'; 
pc_sep = '\'; 
 
if isunix && str(1)==pc_sep 
    str = strrep(str, pc_sep, unix_sep); 
elseif ispc && str(1)==unix_sep 
    str = strrep(str, unix_sep, pc_sep); 
end 
 
if ~isempty(str) && (str(end) == unix_sep || str(end) == pc_sep) 
str(end) = ''; 
end  
 
_______________________________________________________ 
function str = concatpath(varargin) 
unix_sep = '/'; 
pc_sep = '\'; 
 
str = ''; 
for n=1:nargin 
curStr = varargin{n}; 
str = fullfile(str, chompsep(curStr)); 
end 
 
if isunix && str(1)==pc_sep 
    str = strrep(str, pc_sep, unix_sep); 
elseif ispc && str(1)==unix_sep 
    str = strrep(str, unix_sep, pc_sep); 
end 

[Andrew Scott]

 

[Moody]

I've been using multicore for a while now and its absolutely excellent. I'm running on 5 dual xeon x5460 as well as a couple of quad core boxes.  
 
I was wondering if anyone compared performance of this toolbox with the parfor parallel computing matlab toolbox. Are they comparable?  
 
I believe I'm bottlenecked now due to the hard disk I/O, so I was looking at the in memory possibilities of this or potentially upgrading my hd's to solid state to reduce the overhead.  
 
BTW, I also tried precreating all the mat files once instead of doing multiple loops to reduce the I/O. Unfortunately, that didn't help as much as I hoped.  
 
Don't get me wrong though, this is much, much, much faster than single threading, but as always we need to keep pushing :).

[Markus Buehren]

I have opened a discussion group for the Multicore package on Yahoo. Please join and discuss with other users! 
 
http://groups.yahoo.com/group/multicore_for_matlab/

[Richard Crozier]

Fantastic program, and particularly suited to my work with genetic algorithms. There is one mnor error I've noticed though. In startmulticoreslave, if you activate debug mode you reach the following line (77):  
 
fileNr = str2double(regexptokens(parameterFileName,...  
'parameters_\d+_(\d+)\.mat'));  
 
But there is no regexptokens function, at least there isn't in R2007a or R2007b or R2008a. My solution is to replace this with the following lines (although I'm sure someone could come up with something more robust and/or elegent).  
 
fileNrCell = regexp(parameterFileName,'parameters_\d+_(\d+)\.mat', 'tokens');  
 
fileNr = str2double(fileNrCell{1});  
 
The program is excellent though, thanks again!

[Richard]

Brilliant tool! Great being able to sit back and watch a progress bar sliding along as a room full of computers gets to work doing your simulation. 
 
I have a question though - what dimensions are typically used for parameterCell? I've tried doing some large multidimensional runs (e.g. 4x150x10) and things seemed to grind to a halt - I'm still looking into it, just wondered if the dimensions I'm using are typical or too large. 
 
Cheers. Great work, 
Richard.

[dpb10]

 

[Thomas]

Great tool!