Using parfor to parallelyze finite difference scheme
3 views (last 30 days)
I have a rather complex finited differnece scheme im solving with an explicit method (i.e. all values needed to calcualte the next timesteps value are known) and I wanted to parallelize in order to understand how parfor works. When I use parfor my code runs excedingly slowly. I know its not ideal to have parfor inside a for loop (as I now have the pool setup time in each iteration) but I'm not sure how to avoid this since I cannot parallelize the time loop.
When I run the code with Number_Procs = 0, it computes relatively quickly. However, when Num_Procs is anything greater the code is greatly slowed.
I guess I have two main questions:
- Am I using parfor correctly or is there something that I am doing wrong that is causing the slow down?
- Is there a better way to parallilize the inner fro loop to avoid the pool setup time in each iteration?
Notes on scale of the problem:
The Time and Domain vecotrs can really be any size. Currently, I have the time vector of length 10,000 and the Domain vector of length 1,000, just for testing purposes. When I run with Num_procs = 0, the code completed in ~6 seconds. However, when I run with Number_Procs = 4, the code completes in ~185 seconds.
for t = 1:length(Time_vector)
% some setup stuff
parfor z = (1:length(Domain), Number_Procs)
U_ipo = Explicit_Solver(U_i); % Solve for the next timestep using the current
Prudhvi Peddagoni on 28 Dec 2020
parfor doesn't always speeds up the execution time. Sometimes the parallel overhead will be so big that it outweighs the advantage brought in by parallel execution. You can find more information about it here.
Hope this Helps.