Function in parfor loop produces incorrect output, in normal for-loop output is correct

Question

Wouter Grimme 2018-9-5

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/417614-function-in-parfor-loop-produces-incorrect-output-in-normal-for-loop-output-is-correct

回答： Wouter Grimme 2018-9-5

I believe I am being hampered by the knowledge of how a parfor loop work, hopefully someone knowns in which direction to point me.

The problem: I have a script which uses a function to calculate the crash location of an object with {x,y,z} and {vx,vy,vz} coordinates and speeds. I use the simple script below to evaluate the effect of a decreasing time-step dt on the performance and precission. Put in a normal for-loop, the output is as expected. (I understand that I write over my output in the main loop, but ignore that for now).

for dt = [0.1 0.01 0.001 0.0001 0.00001 0.000001]
    for l = 1:10  
        [x(l), y(l)] = f_crash_location(0,0,100,10,0,0,0.0038,dt);
    end
end

The output is then x = 41.244... and y = 0, which is physically correct :). However, in the main simulation where I am using the function, l is not 10, but more like a couple million runs, so I thought, perfect for parallelization. But when I change the for to a parfor-loop like this:

for dt = [0.1 0.01 0.001 0.0001 0.00001 0.000001]
    parfor l = 1:10  
        [x(l), y(l)] = f_crash_location(0,0,100,10,0,0,0.0038,dt);
    end
end

The code is executed insanely fast, which seems nice, but the output is useless. x = 10^-5 is plainly wrong. (I figured out the parfor loop only does dt*x_dot once.)

The code of the function f_crash_location.m is:

 function [x_n,y_n] = f_crash_location(x_n,y_n,z_n,x_dot_n,y_dot_n,z_dot_n,C_D,dt)
    %%Declare variables
    g = 9.807;
    %%Numerical model
    while z_n > 0
        % Vector velocity
        V = sqrt(x_dot_n^2 + y_dot_n^2 + z_dot_n^2);
        % Accelerations
        x_2dot_n = -V*x_dot_n*C_D;
        y_2dot_n = -V*y_dot_n*C_D;    
        z_2dot_n = -g + -V*z_dot_n*C_D;
        % Velocities
        x_dot_n = x_dot_n + x_2dot_n*dt;
        y_dot_n = y_dot_n + y_2dot_n*dt;
        z_dot_n = z_dot_n + z_2dot_n*dt;
        % Positions
        x_n = x_n + x_dot_n*dt;
        y_n = y_n + y_dot_n*dt;
        z_n = z_n + z_dot_n*dt;
    end
end

After playing around a bit I believe it has something to do with how variables are used differently in parfor then in for-loops, but I am stuck on how to proceed. What I am doing wrong that the parfor loop spits out, well, that brown stuff?

4 个评论
显示 2更早的评论隐藏 2更早的评论

Stephen23 2018-9-5

编辑：Stephen23 2018-9-5

在 MATLAB Online 中打开

I cannot replicate the behavior on R2012b using default settings:

for dt = [0.1 0.01 0.001 0.0001 0.00001 0.000001]
    parfor l = 1:10
        [x(l), y(l)] = f_crash_location(0,0,100,10,0,0,0.0038,dt);
    end
    x,y % display output
end

Where parfor prints this:

x =
   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446
y =
     0     0     0     0     0     0     0     0     0     0

And for prints this:

x =
   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873   40.8873
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553   41.2553
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401   41.2401
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444   41.2444
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446
y =
     0     0     0     0     0     0     0     0     0     0
x =
   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446   41.2446
y =
     0     0     0     0     0     0     0     0     0     0

They look the same to me. What MATLAB version are you using?

Stephen23 2018-9-5

编辑：Stephen23 2018-9-5

"Probably I use global variables in a wrong way together with a par-for loop?"

Global variables are the wrong way to do anything.

If you want to know what experienced programmers think about globals, then search the 'net.
If you want to know what experienced MATLAB users think of globals, then search this forum.
If you want to know what the writers of MATLAB think about globals, then read their blogs.
If you want to know why they all think that, then look at your own question: it is a classic example of how globals cause bugs that are hard to track down.

Simple rule to make your life easier and waste less time chasing down hard-to-find bugs: do NOT use global variables. You might find this useful, it describes much better alternatives to global variables:

https://www.mathworks.com/help/matlab/math/parameterizing-functions.html

Wouter Grimme 2018-9-5