Assign 4 GPUs - 1 to each worker in parfor loop

Hi. I have written a function that uses the GPU to speed up the progress. It seems to be working well, because the GPU version is already 7.5 times as fast as the CPU version using 12 workers in a parfor loop. But the server that runs the batch jobs has 4 GPUs. I assume that in my gpu function, only one of them is used. I was wondering if I could write a script that uses a parfor loop with 4 workers (dividing the matrices into 4 segments) and assigns one GPU to each of the workers. Ideally, this should give the script another 4x boost, bringing the total speed boost to 30.

Here is my idea:

fr1 = 4; %Number of GPUs to work
width = size(input1,2);
fr2 = round(width./fr1); %index to divide data by to spread it equally over 4 workers
output1 = zeros(size(input1),'single');
output2 = output1;
output3 = output1;
out1 = cell(1,fr1); out2 = out1; out3 = out1; %store data in cells to avoid indexing errors
parfor fr = 1:fr1 %divide between 4 workers
      [out1{fr}, out2{fr}, out3{fr}] = ...
          GPU_function(input1(: , fr*fr2-fr2+1:min(fr.*fr2+fr2,width)),...
         input2(:,fr*fr2-fr2+1:min(fr.*fr2+fr2,width)),...
         input3 (:,fr*fr2-fr2+1:min(fr.*fr2+fr2,width)));
end
%sort data from cells into outputs
for fr = 1:fr1
  output1(:,fr*fr2-fr2+1:min(fr.*fr2+fr2,width)) = out1{fr};
  output2(:,fr*fr2-fr2+1:min(fr.*fr2+fr2,width)) = out2{fr};
  output3(:,fr*fr2-fr2+1:min(fr.*fr2+fr2,width)) = out3{fr};
end

I am pretty sure this is not enough to assign a GPU to each worker. Running the script on the server results in the error message "Out of memory on device." Running GPU_function() on the entire data set on one worker without splitting it into 4 does not return this error. Does anyone have an idea how I could make this work?

Thanks in advance!

Marc

Answer 1

[Walter Roberson]

http://blogs.mathworks.com/loren/2013/06/24/running-monte-carlo-simulations-on-multiple-gpus/

[Marc Jakobi]

Thanks! I was hoping to find an article like this.