An execution model for adaptive load-balancing on multicore and multi-GPU systems
ET International, Inc., Newark, DE 19711, USA
ETI Technical Memo 2, 2014
@article{lauderdale2014execution,
title={An execution model for adaptive load-balancing on multicore and multi-GPU systems},
author={Lauderdale, Christopher and Springer, Robert and Khan, Rishi},
year={2014}
}
As computing systems increase in size and parallelism, it becomes more and more difficult to balance workload during program execution. Heterogeneous systems further complicate the situation, as their different constituent compute resources may consume work at different rates, and may have affinity for different kinds of work. Traditional approaches to load-balancing fail to fully address the problem, resulting in distributed work but unbalanced load, and may require extensive hand-tuning to result in an efficient execution profile. SWARM (SWift Adaptive Runtime Machine) is a dynamic, event-driven, task-based runtime system that helps solve the problem of load balancing by dynamically binding appropriate work to available resources; it has been extended to incorporate GPUs into its work-scheduling model. This paper compares the performance of a Cholesky decomposition benchmark implemented for SWARM against other industry-standard approaches, OpenMP (CPUonly) and MAGMA (CPU+GPU), on AMD and Intel/Nvidia platforms. We show that SWARM gives performance advantages over these alternatives.
July 20, 2014 by hgpu
