Strongly scalable algorithms for matrix and tensor computations
Title: |
Strongly scalable algorithms for matrix and tensor computations |
DNr: |
SNIC 2015/1-343 |
Project Type: |
SNIC Medium Compute |
Principal Investigator: |
Lars Karlsson <larsk@cs.umu.se> |
Affiliation: |
Umeå universitet |
Duration: |
2015-10-01 – 2016-10-01 |
Classification: |
10105 |
Keywords: |
|
Abstract
The number of processor cores available to researchers and engineers
is rapidly growing. To effectively utilize this increased processing
power, codes must be parallel and scalable. However, many existing
codes are scalable only in the weak sense, that is, they maintain
efficiency when scaling to a larger parallel system only if also the
problem size is scaled up. We are primarily interested in dense matrix
and tensor algorithms and here it is evident that very large problems
are necessary to obtain high efficiency using standard
algorithms. Much of the current research in the area focus on
developing algorithms that are scalable in the strong sense, that is,
algorithms that maintain high efficiency when the problem size remains
fixed. We propose to continue the development of strongly scalable
matrix and tensor algorithms as well as constructing generic
techniques that facilitate the development of such algorithms. For
example, we intend to develop algorithms and software for dense
eigenvalue problems, low-rank tensor computations, and two-sided
matrix decompositions.