Parfor or parfeval, what is better?

if I have some code that could be writed in same way with parfeval o parfor, what i have to use? What is better? Which is more advantageous, faster... e.g:

parfor i = 1:n
  result = function(data);
end

or

for i = 1:n
  F(i) = parfeval(@function, 1, data);
end
for i = 1:n
  [index, value] = fetchNext(F);
end

[Rik]

I couldn't find any comparisons with a quick search, although I did find that parfeval is (or used to be?) 'plagued by memory a leak'. I checked the release notes for mentions of parfeval, but nothing popped up.

In general I would assume that parfor is a better choice, but that might be a knee jerk reaction because of the feval part of its name.

[Edric Ellis]

The memory leak described in that bug report is shown as being fixed in R2015a.

Answer 1

[Edric Ellis]

At least some of the trade-offs are:

• parfor is generally easier to use, and the code probably looks much more like your serial code
• parfor loops can be used by people who don't have Parallel Computing Toolbox
• parfor loops perform a degree of load-balancing to try and minimise overheads
• parfeval gives you complete control over how the work is chunked up for the workers
• parfeval is asynchronous, and lets the MATLAB client get on with other stuff while the workers are busy (e.g. updating plots or other visualisations)