Multiple Parallel Jobs on Server

2 views (last 30 days)
Shumao Zhang
Shumao Zhang on 10 Feb 2022
Answered: Edric Ellis on 11 Feb 2022
Let's say I have a script "main.m" which uses the parallel computing toolbox for parallelization. I have a server (Mac) with 20+ cores of cpus, and would like to run multiple "main.m" jobs at the same time for parameter searching. Each "main.m" job only needs 4 cpus. I opened multiple instances of Matlab and ran one job in each instance. However, I found that the speed of each job slows done significantly (compared to just run one single job). I guess it might because some cores are used by more than one job. Suppose 4*number of jobs < 20, what would be a suggested way to achieve the goal? Is there a way to ensure that each job will use different cores?
As a remark, "main.m" is a script and is not a function.
Shumao Zhang
Shumao Zhang on 10 Feb 2022
Thanks for the reply. However, for each job I only need 4 cores so the size of my parallel pool is 4.
I need to, for example, run two jobs at one time, which means ideally I just need 8 cores and assign them to these two jobs without sharing.
When I tried openning two Matlab instances, I feel that some cores are shared by two jobs so the speed for each job slows down significantly.

Sign in to comment.

Answers (1)

Edric Ellis
Edric Ellis on 11 Feb 2022
You say that "main.m" script uses 4 cores to do parallel computing. I presume this means that it contains a parfor loop. One way you could launch multiple different copies of "main.m" is by using the batch command. Each time you call batch, you get back a parallel.Job object which lets you see how the computation is going, and retrieve the results at the end. By using the 'Pool' parameter to the batch command, you can make the job have a parallel pool available for running the parfor loops. You might do something like this:
for idx = 1:3
param = idx; % param is a variable read by your script "main.m"
job(idx) = batch('main', 'Pool', 4); % Launches "main.m" with a parallel pool of 4 workers
% Now that all jobs are launched, we can wait for them to complete and
% collect the result
for idx = 1:3
results{idx} = load(job(idx)); % This will retrieve variables created by "main.m"


Find more on Parallel Computing Toolbox in Help Center and File Exchange

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by