Multithreaded and PCT sparse matrix inversions?
in [Matlab]
|
Prev: TCP/IP connection lost
Next: why does Matlab fail to release memory even if I clear all the variables?
From: Tim on 1 Mar 2010 08:34 Since this topic will concern my work for the next year or so (and for many other people I'm sure), I'm just wondering if there's a timeline at the Mathworks for the implementation of these two (related but definitely not the same) functions? Specifically: 1) Multithreaded sparse inversion (slash, not LU or QR or etc...) AFAIK a lot of the sparse inversion code is not multithreaded nicely, even as late as 2009b (I just ran some test one again and saw only one core was being used, versus full utilization for full invserions). Now I know there's a lot of shuffling under the hood recently for sparse inversion so pardon me if the latest Matis, SuiteSparseQR, etc are all in the pipeline. It'll be nice to have some idea of where things are now though. Promises are great but just knowing that wrenches are being broken out is awesome as well. 2) Inversion for (co)distributed PCT matrices There's quite a flurry of recent developments in this field as well (i.e. MUMPS). Are there ever gonna be plans to work this into future version of the PC toolbox? I can see how robustness is a issue and a wrong matrix at the wrong time could easily bring down the whole computing network infrastructure, but is this even being considered? It'll be great to know, even if the error message you get when trying to invert a sparse distributed matrix says "not YET implemented" (wink).
From: Bruno Luong on 1 Mar 2010 10:22
"Tim " <evitaerc(a)gmail.com> wrote in message <hmgfot$asu$1(a)fred.mathworks.com>... > Now I know there's a lot of shuffling under the hood recently for sparse inversion so pardon me if the latest Matis, SuiteSparseQR, etc are all in the pipeline. Sparse QR is already implemented in the few most recent releases. Bruno |