Enable cross-iteration parallelism for PIM-based graph processing with vertex-level synchronization

Abstract

Processing-in-memory (PIM) architectures have emerged as a promising solution for accelerating graph processing by enabling computation in memory and minimizing data movement. However, most existing PIM-based graph processing systems rely on the Bulk Synchronous Parallel (BSP) model, which frequently enforces global barriers that limit cross-iteration computational parallelism and introduce significant synchronization and communication overheads.

To address these limitations, we propose the Cross Iteration Parallel (CIP) model, a novel vertex-level synchronization approach that eliminates global barriers by independently tracking the synchronization states of vertices. The CIP model enables concurrent execution across iterations, enhancing computational parallelism, overlapping communication and computation, improving core utilization, and increasing resilience to workload imbalance. We implement the CIP model in a PIM-based graph processing system, GraphDF, which features a few specially designed function units to support vertex-level synchronization. Evaluated on a PyMTL3-based cycle-accurate simulator using four real-world graphs and four graph algorithms, CIP running on GraphDF achieves an average speedup of 1.8$\times$ and a maximum of 2.3$\times$ compared to Dalorex, the state-of-the-art PIM-based graph processing system.