Event-Triggered Communication in Parallel Computing
Authors: Soumyadip Ghosh (University of Notre Dame)
Abstract: Communication overhead in parallel systems can be a significant bottleneck in scaling up parallel computation. In this paper, we propose event-triggered communication methods to reduce such communication overhead for numerical simulation of partial differential equations. As opposed to traditional communication in which processing elements exchange data at every iteration of the numerical algorithm, the main idea behind event-triggered communication is to exchange data only when necessary as dictated by a suitably defined event. We show through numerical experiments that these methods have potential to reduce simulation time. Strong scaling plots show that the approach may be effective on large machines as well.
Back to 9th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems Archive Listing
Back to Full Workshop Archive Listing