Research Projects
Communication-Avoiding Krylov Subspace Methods
Kaushik Datta, Mark Frederick Hoemmen, Marghoob Mohiyuddin, Laura Grigori1, James Demmel and Katherine A. Yelick
We are currently working on modifying Krylov subspace methods to exploit techniques for avoiding communication when computing sparse matrix-vector products and QR factorizations. Open problems include developing compatible preconditioners and proving stability, as well as the computational challenges of implementing the kernels efficiently.
1IRISA
More information: http://bebop.cs.berkeley.edu