New Algorithms/Software
Symmetric ?-problem
- Divide and Conquer (J. Cuppen, J. Dongarra, D. Sorensen, L. Jessup, F. Tisseur, ...)
- recursively split problem
- generates lots of parallel work
- Holy Grail (J. Demmel, B. Parlett, I. Dhillon,...)
- Optimal output complexity O(n2) and embarrassingly parallel
- Works by redefining clusters of eigenvalues to get relative gap instead of absolute gap.
-
However, Devil’s in the details in terms of implementation