Project abstract for group karypisg

High-Performance and Big Data Research

This group's research during 2015 focused on the development of parallel shared-memory graph partitioning, ordering, and clustering algorithms that use the multilevel paradigm. Graph partitioning is used widely for parallel task scheduling and data distribution. Graph ordering is used reducing the amount of computation and memory required for sparse direct numerical methods. Graph clustering is a widely used technique for discovering relationships between data points by creating groups of unconstrained size with high internal connectivity. Access to MSI's HPC resources has been critical in the development of these algorithms as evaluating the scalability of the algorithms requires machines with a large number of compute cores, and many of the graphs/matrices in these domains reach massive size, requiring large amounts of memory.

The group's work in 2016 focuses on developing hybrid shared/distributed memory codes that can effectively utilize compute architectures composed of many multicore nodes. This work will be an extension of the researchers' past work on shared and distributed memory graph partitioning, ordering, and clustering. Part of this will include ensuring their methods scale to very large numbers of processing cores. These methods will be required for partitioning, ordering, and clustering problems on the next generation of large petascale and exascale machines, which will have millions of processing cores.

Return to this PI's main page.