P86: HyGraph: High Performance Graph Processing on Hybrid CPU+GPUs platforms
Abstract: Graph analytics is becoming increasingly important in many domains, such as in biology, social sciences, and data mining. Many large-scale graph-processing systems have been proposed, either targeting distributed clusters or GPU-based accelerated platforms. However, little research exists on designing systems for hybrid CPU-GPU platforms, i.e., exploiting both the CPU and the GPU efficiently.
In this work, we present HyGraph, a novel graph-processing system for hybrid platforms which delivers performance by using both the CPU and GPUs concurrently. Its core feature is dynamic scheduling of tasks onto both the CPU and the GPUs, thus providing load balancing, contrary to the state-of-the-art approach based on static partitioning. Additionally, communication overhead is minimized by overlapping computation and communication.
Our results demonstrate that HyGraph outperforms CPU-only and GPU-only solutions, delivering close-to-optimal performance. Moreover, it supports large-scale graphs which do not fit into GPU memory and is competitive against state-of-the-art systems.
Award: Best Poster Finalist (BP): no
Two-page extended abstract: pdf