I am rally confused. Why parfor is slower than for in this very simple test situation?

1 次查看（过去 30 天）

Mr M. 2018-1-5

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/375739-i-am-rally-confused-why-parfor-is-slower-than-for-in-this-very-simple-test-situation

编辑： Adam 2018-1-5

tic
A = 1:100;
B = 1:100;
Nx = 500;
Ny = 500;
x_values = linspace(0,1,Nx);
y_values = linspace(0,1,Ny); 
collected_data = zeros(Nx,Ny);
parfor ix = 1:Nx
    x = x_values(ix);
    for iy = 1:Ny
        y = y_values(iy);
        collected_data(ix,iy) = exp(sum(A*x+B));
    end
end
toc

Elapsed time is 1.121132 seconds. (with 2 workers)

But: Elapsed time is 0.149634 seconds. (with for instead of parfor)

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Adam 2018-1-5

编辑：Adam 2018-1-5

Parallelisation can be a bit of a black art when it comes to working out how to optimise it and when it is worth using. There is an overhead to copying data to and from workers. If the computation being done is not sufficiently complex to balance that out then the data transfer time will outweigh the computation time.

Coming up with a 'simple test situation' for parfor, therefore, is not always trivial as it cannot be too simple if you want to compare timings.

Also, in your case, when I put that code into a file I get a warning about y_values being a broadcast variable. You shouldn't ignore warnings if you care about speed!

请先登录，再进行评论。

请先登录，再回答此问题。