Avoid broadcasting an array in parfor

2 vues (au cours des 30 derniers jours)
dipak sanap
dipak sanap le 20 Fév 2022
Commenté : dipak sanap le 20 Fév 2022
Hi, I have a matrix (A) and I want to access certain columns of this matrix in for loop. I need some help avoiding the broadcasting the the entire matrix A in the loop when I use parfor. Any help would be useful. I have provided simple working example which is a part of larger code.
A = rand(100,5);
combinations = randi([1 5],100,3);
parfor i=size(combinations,1)
current_comb = combinations(i,:);
current_profile = A(:,current_comb);
end

Réponse acceptée

Raymond Norris
Raymond Norris le 20 Fév 2022
As a side note, I suspect that
parfor i=size(combinations,1)
should be more like
parfor i=1:size(combinations,1)
as the original is a scalar and therefore, the for-loop is a single iteration.
In your example, combinations is a sliced variable; however, A needs to be a broadcast. Using the loop variable, i, MATLAB knows a prior how to break up combinations to each of the workers. On the other hand, indexing into A is not known until we know the value of current_comb (at runtime). To provide this flexibleness, MATLAB needs to send all of A as a broadcast variable.
It might help to look at ticBytes and tocBytes to get a sense of how much data is being sent back and forth.
  3 commentaires
Raymond Norris
Raymond Norris le 20 Fév 2022
That's the right idea. Now you're able to slice A_combs. The only caveat is that you've doubled your memory size on the client side (with A_combs). Mind the typo in cuurent_profile.
One consideration is how much of this can be created within the parfor loop. You've given pseudo code, but if you're really using rand, you might generate your combinations in the parfor, eliminating the need to send it to the workers.
dipak sanap
dipak sanap le 20 Fév 2022
Hi, thanks Raymond. Yeah, it was just a pesudo code. I dont actully use rand to generate combinations and they can not be done in par loop. But, thanks I just wanted to make sure that creating A_combs was correct approch. I would go ahead with this strategy.

Connectez-vous pour commenter.

Plus de réponses (0)

Catégories

En savoir plus sur Parallel for-Loops (parfor) dans Help Center et File Exchange

Community Treasure Hunt

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

Start Hunting!

Translated by