On 20111112@18:25, Jordi Gutiérrez Hermoso wrote: > On 12 November 2011 18:18, Carlo de Falco <carlo.defa...@gmail.com> wrote: > > On 12 Nov 2011, at 16:04, c. wrote: > >> On 12 Nov 2011, at 12:17, Alexander Barth wrote: > > .... > > It looks like it's working: > > > >>> n = 1000; > >>> A = 15 * eye (n) + sprandn (n, n, .2); > >>> b = ones (n, 1); > >>> P = diag (diag (A)); > >>> tic, [x, flag] = gmres (A, b, [], 1e-7, n, P); toc > > Elapsed time is 1.73299 seconds. > >>> As = sparsersb (A); > >>> tic, [x, flag] = gmres (As, b, [], 1e-7, n, P); toc > > Elapsed time is 1.40307 seconds. > > > > the speed-up is nothing to write home about though but, still, > > it's a really nice package and already usable, > > so Michele keep up with the good work :) > > That's a tiny sparse matrix. I'm curious about how the algorithm > scales. Can you try a couple of orders of magnitude larger? > > Also, what version of Octave is that? Are you working on dev or stable? Hahahaha :) Yes Jordi is right: you should try much large matrices, possibly with tens of elements per row. See the articles on http://claudius.ce.uniroma2.it/~martone/ as soon as the webserver is back to life. I have yet to document much of librsb: there are different tuning parameters involved in librsb's operation.
But generally, with smaller matrices (e.g.: as a rule of thumb, ones fitting in the L2 cache) it's not even said that operation was really parallel (it depends on the matrix partitioning; maybe this could be interesting to "visualize" in some way...; if the "recursive partitioning" was not applied at all, operation was serial).
pgp7fJmwlHGxC.pgp
Description: PGP signature
------------------------------------------------------------------------------ RSA(R) Conference 2012 Save $700 by Nov 18 Register now http://p.sf.net/sfu/rsa-sfdev2dev1
_______________________________________________ Octave-dev mailing list Octave-dev@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/octave-dev
