Sparse data and irregular problems in scientific applications

10 Mar 2021
16:10 - 16:30

Sparse data and irregular problems in scientific applications

The dominant paradigm in scientific computation is built around “dense” data, which is assumed to be highly regular and well behaved. We explore the tensions with this ideal situation, by looking at several sparsity generating mechanisms, as well as examples of naturally sparse data such as in graphs or long-tailed distributions. Sparsity in Machine Learning is also considered. We briefly describe how sparsity impacts performance on standard architectures, and we touch upon the migrating threads architecture as a candidate for improving performance on irregular problems.