spmd, parfor, createParellel, Slower than the usual code without going in parallel

Question

Puiyin 2014-1-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/111366-spmd-parfor-createparellel-slower-than-the-usual-code-without-going-in-parallel

评论： Edric Ellis 2014-1-6

I have the following code which is quite computationally intensive. I tried to make it go in parallel. It turned out the computation time is much slower than that without going in parallel. Does anyone know what is wrong with my code? I have the following 2 examples that both showed this weird result. I tried replacing parfor with spmd and ran it. It produced the weird result too. Thank you very much.

1st:

% Conditional_pdf is a probability function that takes vectors as inputs and output 
%a vector of elements as probability corresponding to each vector element. 
% x_t is a vector
% x_t1 is a vector
tic;
F = @(x)Conditional_pdf(dt,x_t,x,sigma,beta);
parfor i = 1:length(x_t1)
      ret(i) = mean(F(x_t1(i)));
end
toc;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2nd:

parfor pt=1:2
    f0 = Fwd;
    F = ones(N(pt),M)*f0(1,pt);
    V = ones(N(pt),M)*ALPHA(pt);
      for j=1:N(pt)
          F(j+1,:) = F(j,:) + V(j,:).*abs(F(j,:)).^(BETA(j)).*dW_first(j,:).*sqrt(dt);
          V(j+1,:) = V(j,:) + LAMBDA(j)*(LAMBDA_AVG(j) - V(j,:)).*dt + NU(j).*V(j,:).*dW_second(j,:).*sqrt(dt);
      end
      K = Strike(:,pt);
      Fwd_Premium_call = zeros(length(K),1);
      Fwd_Premium_put = zeros(length(K),1);
      for k=1:length(K)
          temp=sum(max(F(N(pt)+1,:)-Strike(k,pt),0));
          temp1=sum(max(Strike(k,pt)-F(N(pt)+1,:),0));
          Fwd_Premium_call(k) = temp/(M);
          Fwd_Premium_put(k) = temp1/(M);
      end
      Price_call = exp(-LIBOR_d(:,pt).*T_future(:,pt)).*Fwd_Premium_call;
      Price_put = exp(-LIBOR_d(:,pt).*T_future(:,pt)).*Fwd_Premium_put;
      model_volatility_call = blkimpv( Fwd(:,pt), Strike(:,pt), LIBOR_d(:,pt), T_future(:,pt) - T_now(:,pt), Price_call, 10, 1e-10,{'call'});
      model_volatility_put = blkimpv( Fwd(:,pt), Strike(:,pt), LIBOR_d(:,pt), T_future(:,pt) - T_now(:,pt), Price_put, 10, 1e-10,{'put'});
      SABR_vol(:,pt) = [model_volatility_put(1); model_volatility_put(2); model_volatility_call(3); model_volatility_call(4); model_volatility_call(5)];
end

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Edric Ellis 2014-1-6

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/111366-spmd-parfor-createparellel-slower-than-the-usual-code-without-going-in-parallel#answer_120122

There are several reasons why your code might not speed up when going parallel. The most likely is that you're using local workers only, and MATLAB is already doing a good job of multithreading your code - in which case, PARFOR and SPMD cannot speed things up for you. Check using Windows Task Manager, or UNIX "top" to see whether the non-PARFOR version of your code is able to run multithreaded.