parfeval evaluation is slow

Hi

I am using parfeval to calculate the eigenvalues of a matrix M with typical dimension of 1500x1500x100. The parfeval evaluations is slow since I read the output just after the evaluation of each matrix M(:,:,k) k=1,..,100. I want to store all the eigenvalues and eigenvectores in the variabel f and read the output after finishing the parfeval evaluation.

Current code:

eigMtr=complex(zeros(2*L,2*NumOfSlabs*L));
W=eigMtr;
W_1=complex(zeros(L,2*NumOfSlabs*L));W_2=W_1;
for k=1:NumOfSlabs
    f=parfeval(@eig,3,M(:,:,k));
    [Wt,Dt] = fetchOutputs(f);
    start=1+(k-1)*2*L;stop=2*k*L;
    eigMtr(1:2*L,start:stop)=sqrt(Dt);
    W1=Wt(1:L,:); 
    W2=Wt(L+1:2*L,:);
    W_1(:,start:stop)=W1;
    W_2(:,start:stop)=W2;
    W(1:L,start:stop)=W2;
    W(L+1:2*L,start:stop)=W1; 
end

I want to do something like this

for k=1:NumOfSlabs
    f=parfeval(@eig,3,M(:,:,k));
end

for k=1:NumOfSlabs
    [idx,Wt,Dt] = fetchNext(f)
    start=1+(k-1)*2*L;stop=2*k*L;
    eigMtr(1:2*L,start:stop)=sqrt(Dt);
    W1=Wt(1:L,:); 
    W2=Wt(L+1:2*L,:);
    W_1(:,start:stop)=W1;
    W_2(:,start:stop)=W2;
    W(1:L,start:stop)=W2;
    W(L+1:2*L,start:stop)=W1; 
end

Thanks for your help

Poul-Erik

Answer 1

[Walter Roberson]

Note the correction to the number of output arguements parameter in the parfeval() call

f = cell(NumOfSlabs, 1);
for k=1:NumOfSlabs
    f{k} = parfeval(@eig,2,M(:,:,k));
end

wait(f{1});

for k=1:NumOfSlabs
    [Wt,Dt] = fetchOutputs(f{k});
    start=1+(k-1)*2*L; stop=2*k*L;
    eigMtr(1:2*L,start:stop)=sqrt(Dt);
    W1=Wt(1:L,:); 
    W2=Wt(L+1:2*L,:);
    W_1(:,start:stop)=W1;
    W_2(:,start:stop)=W2;
    W(1:L,start:stop)=W2;
    W(L+1:2*L,start:stop)=W1; 
end

[Poul-Erik Hansen]

Thanks for the answer. How can you be sure that fetchOutputs preserve the calculation order (k=1,..., NumOfSlabs)?. Parfeval performes the calculations asynchronous. Isn't it better to uses fetchNext?

FethNext returns the evaluation index, idx, that make it possible to write

start=1+(idx-1)*2*L

stop=2*idx*L

....

[Walter Roberson]

Each parfeval() generates a singular future. We write those futures in cell arrays, so when we reference the cell array at a particular location, it is the location of a single future.

I did things this way to duck the question of whether futures could be stored using standard array indexing, which was something I could not find readily documented. However, looking at https://www.mathworks.com/help/matlab/ref/parallel.future.aftereach.html#mw_2abc4dc5-73a6-4a2b-8d9f-c6f09b39f6d5 I see that you can use regular indexing to create an array of futures. So perhaps the flow could be

for k=1:NumOfSlabs
    f(k) = parfeval(@eig,3,M(:,:,k));
end

for k=1:NumOfSlabs
    [idx,Wt,Dt] = fetchNext(f);
    start = 1+(idx-1)*2*L; stop=2*idx*L;
    eigMtr(1:2*L,start:stop)=sqrt(Dt);
    W1=Wt(1:L,:); 
    W2=Wt(L+1:2*L,:);
    W_1(:,start:stop)=W1;
    W_2(:,start:stop)=W2;
    W(1:L,start:stop)=W2;
    W(L+1:2*L,start:stop)=W1; 
end

[Poul-Erik Hansen]

Exactely. Thank you :-)

[Edric Ellis]

@Walter Roberson It is definitely expected that you can create arrays of parallel.Future - both fetchNext and fetchOutputs accept arrays of futures (in particular, fetchNext is only really useful with an array of futures).

[Walter Roberson]

@Edric Ellis

The documentation for parfeval() does not show any arrays of futures, so I was under the impression that arrays of futures were potentially only generatable by parfevalOnAll

[Edric Ellis]

@Walter Roberson thanks for that - I'll see if we can get an example showing the expected usage on the parfeval page itself. Your comment above has exactly the code I would expect to see.

(A single call to parfevalOnAll returns a scalar parallel.FevalOnAllFuture - and while you can make arrays of these, it isn't as useful, and you can't use those with fetchNext)

[Edric Ellis]

Ah, actually, the Parallel Computing Toolbox page for parfeval already does have examples showing arrays https://uk.mathworks.com/help/parallel-computing/parallel.pool.parfeval.html . @Walter Roberson were you looking at the page in the MATLAB doc? https://uk.mathworks.com/help/matlab/ref/parfeval.html .

[Walter Roberson]

You are right, I was looking at https://uk.mathworks.com/help/matlab/ref/parfeval.html

Answer 2

[Steven Lord]

Have you compared the time of your parfeval call to the time required for a call to the pageeig function?

[Poul-Erik Hansen]

Yes, the two functions have similar speed but different cpuload.

pageeig cpu load 50%

parfeval cpuload 100 %