Fine-Grained Task Migration for Graph Algorithms Using Processing in Memory

Graphs are used in a wide variety of application domains, from social science to machine learning. Graph algorithms present large numbers of irregular accesses with little data reuse to amortize the high cost of memory accesses, requiring high memory bandwidth. Processing in memory (PIM) implemented through 3D die-stacking can deliver this high memory bandwidth. In a system with multiple memory modules with PIM, the in-memory compute logic has low latency and high bandwidth access to its local memory, while accesses to remote memory introduce high latency and energy consumption. Ideally, in such a system, computation and data are partitioned among the PIM devices to maximize data locality. But the irregular memory access patterns present in graph applications make it difficult to guarantee that the computation in each PIM device will only access its local data. A large number of remote memory accesses can negate the benefits of using PIM. In this paper, we examine the feasibility and potential of fine-grained work migration to reduce remote data accesses in systems with multiple PIM devices. First, we propose a data-driven implementation of our study algorithms: breadth-first search (BFS), single source shortest path (SSSP) and betweenness centrality (BC) where each PIM has a queue where the vertices that it needs to process are held. New vertices that need to be processed are enqueued at the PIM device co-located with the memory that stores those vertices. Second, we propose hardware support that takes advantage of PIM to implement highly efficient queues that improve the performance of the queuing framework by up to 16.7%. Third, we develop a timing model for the queueing framework to explore the benefits of work migration vs. remote memory accesses. And, finally, our analysis using the above framework shows that naïve task migration can lead to performance degradations and identifies trade-offs among data locality, redundant computation, and load balance among PIM devices that must be taken into account to realize the potential benefits of fine-grain task migration.