Trying to use 'parfor', but it is very slow

Question

Lee Hwihyeong 2015-2-20

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/179468-trying-to-use-parfor-but-it-is-very-slow

评论： Adam 2015-2-20

Hello, I am trying to use parfor as below.

-------------------------

parpool(2)
A=zeros(10,10);
for iX=1:10
      parfor iY=1:10
          A(iY,:)=rand(1,10);
      end
end

----------------------

However, the case of replacing the 'parfor' with 'for' is 20 times faster. Above code is simplified version of what I'm using. In my real case, 'parfor' should be inside 'for'. Is this a problem?

Could anyone explain why?

Thank you in advance.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Adam 2015-2-20

As per Doug's answer since you also only have 2 cores (or are only using 2 cores) you will have to be doing more work to make the parfor setup worthwhile than if you had e.g. 8 cores because the absolute maximum theoretical speedup factor will only be 2 and realistically it is obviously lower than that even when running optimally.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Doug 2015-2-20

3
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/179468-trying-to-use-parfor-but-it-is-very-slow#answer_168713

You're not doing enough work in the parfor loop. Since there is an overhead associated with parfor, you have to perform enough operations in the loop to make it worthwhile. Also, it's better if your parfor is not inside a for loop, since otherwise you incur the parfor startup overhead each iteration. You may want to try an example such as http://www.mathworks.com/help/distcomp/examples/simple-benchmarking-of-parfor-using-blackjack.html .