Work Stealing in a Shared Virtual-Memory Heterogeneous Environment A Case Study with Betweenness Centrality

This paper uses betweenness centrality as a case study to research efficient work stealing in a heterogeneous system environment. Betweenness centrality is an important algorithm in graph processing. It presents multiple-level parallelism and is an interesting problem to exploit various optimizations. We investigate queue-based work stealing to distribute its tasks across GPU compute units (CUs) and across the CPU and the GPU, which has not been done by prior work. In particular, we demonstrate how to leverage the new platform-atomic operations on AMD Accelerated Processing Units (APUs) to operate cross-device queues in a lock-free manner in shared virtual memory. To make the work stealing runtime and the application more efficient, we apply new architectural features, including atomic operations with different memory scopes and orderings for different synchronization scenarios. We implement our solution using heterogeneous system architecture (HSA). Our results show that betweenness centrality with CPU-GPU work stealing achieves an average of 15% (up to 30%) performance improvement over GPU-only execution for diverse graph inputs. Our work stealing solution can be applied widely to other applications too. Finally, we analyze important parameters critical for queuing and stealing.