Reposting from a question I got offline: IterativeSolvers.jl implements a basic GKL SVD, but it has not been tested for performance with distributed arrays. The project I have in mind will consist of benchmarking and rewriting any necessary parts for speed. Most of the work I foresee coming from improving the speed of parallel matrix-vector products, and particularly implementing linear algebra operations for sparse distributed matrices, which don't exist right now.
There are also questions of how to deal with numerical stability issues and reorthogonalization, and how to design an implementation that allows users fine-grained control of reorthogonalization for speed-accuracy tradeoffs. Thanks, Jiahao Chen Research Scientist MIT CSAIL Thanks, Jiahao Chen Research Scientist MIT CSAIL On Thu, May 28, 2015 at 11:43 AM, Jiahao Chen <jia...@mit.edu> wrote: > I'd be happy to mentor someone working on parallel linear algebra. The > simplest thing to do that will have very high impact is to implement high > performance iterative (Golub-Kahan-Lanczos) SVD, similar to what is > implemented in PROPACK. I'm also interested in a randomized SVD version > similar to what is described in Halko, Martinsson and Tropp, > doi:10.1137/090771806. > > I'm sure there are plenty of ODE projects around, but I would like to see > someone take up the implementation of geometric integrators in ODE.jl. > > Thanks, > > Jiahao Chen > Research Scientist > MIT CSAIL >
