Am I running in parallel? (best way to check)

Is there a better way than this?

function answer = is_in_parallel ()
    try
        answer = ~isempty(getCurrentTask());
    catch err
        if ~strcmp(err.identifier, 'MATLAB:UndefinedFunction')
            rethrow(err);
        end
        answer = false;
    end
end

[Ryan G]

I don't understand the question. If you are running code in a parfor loop, it will be distributed to the workers available.

[Daniel Shub]

Why do you need this? What is your definition of "best"?

[Jan]

@Ryan: It might be useful when a subfunction is called from inside a PARFOR loop.

@Daniel: I assume that "best" means: robust, minimal overhead, runs with and without parallel toolbox.

[Felipe G. Nievinski]

@Jan: you're on mark -- thanks for both clarifications.

Parallel workers are headless (w/ no display), so the print function only supports Ghostscript drivers, which ignores the resolution option. This all happens silently. I'm checking is_in_parallel() to defer making and saving figures until the parfor is done, at which point I start a sequential for-loop.

[Matt J]

Why isn't it as simple as issuing the figure save commands after the

 parfor...end

block? PARFOR won't return control to your mfile until it is finished.

[Felipe G. Nievinski]

@Matt: the user's preferences are as follows:

1. If the Parallel Toolbox is unavailable, make & save figures as soon as possible (in the first loop) -- no need for a second loop;
2. If the Parallel Toolbox is available, don't make crappy figures -- do only the calculations in the first parallel loop, then have a second sequential loop just to overcome parfor's limitations.

[Felipe G. Nievinski]

Other possibilities that I considered for checking were:

 ~isempty(ver('distcomp'))

and:

 (matlabpool('size') > 0)

but neither seemed better than:

 ~isempty(getCurrentTask())

[Tom Holden]

getCurrentTask throws an error within a thread based parallel pool. So perhaps answer should be true if an error is thrown!

Answer 1

[Jan]

Let me summarize the current discussion:

No. There seems to be no better solution.

[Carlos M. Velez S.]

parallel = gcp('nocreate');
if isempty(parallel)
    S.parallel = 'off';
else
    S.parallel = ['on (' num2str(parallel.NumWorkers) ' workers, Cluster: ' parallel.Cluster.Profile ')'];
end

Answer 2

[埃博拉酱]

parallel.internal.pool.isPoolThreadWorker||~isempty(getCurrentJob))

This code can robustly return true if and only if run on a parallel worker either as threads or processes.

[Roman Krylov]

Worked for me. Thanks!

Answer 3

[Tom Holden]

~usejava( 'desktop' ) seems to work even on thread based parallel pools.

Answer 4

[Daniel Shub]

MATLAB has a number of such functions that tell you about the state of the world (e.g., ispc, isdeployed, isjava). These functions seem to be necessary since there are thing you can/cannot do when they are true/false. I am not sure that something like isparallel is needed. It seems you are looking for something more like isheadless. That said, I am not sure this is the best approach. I am guessing here, but it sounds like your two uses cases are something like

parfor = 1:N
  myFuncThatComputesAndPrints;
end

and

for = 1:N
  myFuncThatComputesAndPrints;
end

and you want myFuncThatComputesAndPrints to know if it is running headless so that it will suppress the printing. I would suggest that you modify myFuncThatComputesAndPrints to take a flag to determine if it should print giving you

printFlag = true;
parfor = 1:N
  myFuncThatComputesAndPrints(printFlag);
end
myFuncThatOnlyPrints;

and

printFlag = false;
for = 1:N
  myFuncThatComputesAndPrints(printFlag);
end

This would allow you to expand the cases where you choose not to print.

[Jan]

I see benefits for isparallel as well as for isheadless as well as for using a flag to trigger a specific behavior. I recommend not to implement a FOR and a PARFOR version of the code, because maintaining and debugging two versions of the same code tends to result in two different functions, which will diverge faster as expected.

Therefore I would prefer: printFlag = isheadless; or printFlag = isparallel depending of what describes the job more accurate. This has the advantage, that you can test the single-thread mode on a multi-core machine easily.

[Felipe G. Nievinski]

@Jan: again, right on mark.

I'd rather not hard-code the printFlag value -- that's the very raison d'etre for isparallel(). I'd take isheadless(), were it provided by MATLAB.

Thanks for the comments everybody. I just wanted to check whether I was missing something.

The final code looks something like this:

parallel = false(N,1);
parfor i=1:N
  result{i} = myFuncThatComputes();
  parallel(i) = isparallel();
  if parallel(i),  continue;  end
  myFuncThatPrints(result{i});
end
if any(parallel)
  for i=1:N
    myFuncThatPrints(result{i});
  end
end

[Daniel Shub]

My point was not to hard code printFlag, rather it was to move it outside the parfor loop. I would suggest:

printFlag = ~isparallel();
parfor i=1:N
  result{i} = myFuncThatComputes();
  if ~printFlag
    myFuncThatPrints(result{i});
  end
end
if printFlag
  for i=1:N
    myFuncThatPrints(result{i});
  end
end

I think this will be more efficient since you only have to call isparallel once. The key difference, however, is that within this framework it is not really about checking isparallel, but rather if you want to print. At this point, the only reason you may not want to print might be because you are running in parallel. Later, however, you might have a headless workstation. In that case then you could do something like

printFlag = isparallel | isheadless

Later you might want to not print if it is parallel, headless, or someplace that doesn't have a printer.

[Felipe]

How would you implement isparallel that would work if called outside the parfor? I don't think any of ~isempty(ver('distcomp')), (matlabpool('size') > 0), or ~isempty(getCurrentTask()) would work.

[Jan]

getCurrentTask() works outside a PARFOR loop and replies [].