Reducing Communication Costs in the Parallel Algebraic Multigrid
Presenter
Event Type
Doctoral Showcase
Algorithms
Applications
Architectures
Heterogeneous Systems
TimeTuesday, November 14th10am - 5:15pm
LocationFour Seasons Ballroom
DescriptionAlgebraic multigrid is an optimal sparse linear solver that suffers from poor parallel scalability due to large communication requirements. This costly communication can be improved through changes to both the algorithm and parallel implementation. The amount of data required to be communicated can be reduced through sparsification of the matrices, yielding slower convergence but significantly cheaper iterations. Furthermore, the data that must be communicated can be sent between processes with increased efficiency through the use of topology-aware methods. The number of messages injected into the network can be reduced greatly at the cost of additional intra-node communication, yielding cheaper communication on coarse levels of AMG.
