Parallel rank calculation for sparse matrices -- suggestions?
4 views (last 30 days)
Show older comments
Barry on 30 Aug 2022
Commented: Bruno Luong on 6 Sep 2022
I need to calculate the rank of large (> 1 terabyte of non-zero elements) sparse matrices with MATLAB. Exploring the Parallel Toolbox, but can't seem to find anything that convinces me what is offers will be helpful. If I'm wrong, can someone point me in the right direction? Ideally, I'd just want to take my existing code that has "r=sprank(A)" in it and have that library call run in parallel, perhaps with some additional annotation as needed. Would using a GPUARRAY help here, for example? Doesn't seem like it, but perhaps I'm wrong. Hoping someone here can help me out. Thanks!
Professor of Computer Science
Bruno Luong on 6 Sep 2022
"After doing a little digging, I now believe sprank() merely counts the number of non-zero columns. This is not what I require."
It does something more intelligent than that it match the row to column through Dulmange Mendelsohn permutation. In your case might be all the non-zero column can be matched. But that is exactly what structural rank means.
Bruno Luong on 30 Aug 2022
Moved: Bruno Luong on 30 Aug 2022
Edric Ellis on 1 Sep 2022
Yes, the "extended capability" description could probably do with some refinement to make it clear exactly what works. (Today, there are a good number of MATLAB functions that cannot be run on a thread pool worker, and this "extended capability" really means simply that the function can run on a thread pool worker).
The performance situation is ... complicated, unfortunately. There is no single simple answer as to whether running N copies of a given function concurrently on workers is faster than running N copies sequentially on the client. It depends on all sorts of details of the implementation of that function - but primarily whether the function is already intrinsically multithreaded by MATLAB itself.
More Answers (2)
Matt J on 30 Aug 2022
Is the matrix square or is it tall and thin? If the later, then it may be an easier computation to compute rank(A.'*A).
Matt J on 5 Sep 2022
Edited: Matt J on 5 Sep 2022
I would take caution before relying on rref. It has been often found to lack robustness, see e.g.,
Bruno Luong on 5 Sep 2022
For thin and tall sparse matrix A of size (m x n), m>>n and n in the order of 1000s, it might be possible to compute the rank using q-less qr, which is better than rank(A'*A) which has the drawback of square the condition number.
Find more on Sparse Matrices 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!