log() much slower in parallel

I run a MATLAB parallel job using the local scheduler to increase performance of the calculation. However, my function is much slower in the parallel job compared to its serial execution. By profiling I found out that some operations take much more time in parallel tasks (e.g. calling the built-in function log() about 6 times more). Any idea why it is slower?

A simplified example, let us have the following three functions:

function par_test()

   Scheduler = findResource('scheduler', 'type', 'local');
   Job = createJob(Scheduler);

   createTask(Job, @par_func, 1, {});   

   submit(Job);
   waitForState(Job, 'finished') 

   OutputArgCell = getAllOutputArguments(Job);
   pInfoVector = [OutputArgCell{:, 1}];
   mpiprofile('viewer', pInfoVector);

   destroy(Job);
end

function pInfo = par_func()

   mpiInit;
   mpiprofile on

   par_calc()   

   pInfo = mpiprofile('info');
end

function par_calc()

   U = unifrnd(0, 1, 10000, 10000);     
   R = log(U);  
end

Run

>> par_test

to execute par_calc in one parallel task and generate a profile report.

Run

>> profile on

>> par_calc

>> profile viewer

to execute par_calc serially and generate a profile report. Compare the execution time of log().

Answer 1

[Walter Roberson]

log() is, I believe, normally automatically parallized over the number of available cores, even without the Parallel Computing Toolbox. If you have 6 cores then 1/6 of the work (approximately) would go to each core. But then when you run in parallel, each session only has access to a single core and cannot distribute the session's work.