|
Effective partitioning and load balancing
Boman
Wednesday, November 7th
11:00 - 11:45 am
A good data distribution is crucial to achieve good performance in parallel
computing. Partitioning and load balancing are techniques to distribute the
work (load) approximately evenly among processors, while keeping the
communication low. Combinatorial models like graphs and hypergraphs are
powerful tools. We discuss these models in the context of sparse matrix
computations. We show that two-dimensional data distributions can reduce
the communication requirement compared to the commonly used row (or column)
distributions.
We give an overview of Zoltan, a general-purpose toolkit for load
balancing, which will soon become available as a Trilinos package, and
Isorropia, a current Trilinos package that interfaces with
Zoltan.
|