ByteGAP: A Non-continuous Distributed Graph Computing System using Persistent Memory

Graph computing systems play a critical role in a variety of industrial applications. This study examines ByteDance's graph computing system workload, which challenges the conventional notion of a one-shot, lightweight graph computing task that can scale to trillions of edges. The workload includes both small and large-scale tasks separated by a 1000-second runtime threshold. The majority of the workload is dominated by small-scale tasks submitted arbitrarily, but with high time-sensitive requirements. Large-scale tasks make up the bulk of computing resources and occur periodically. Therefore, the graph computing system must be capable of pausing running tasks and prioritizing more critical ones. In this paper, we introduce ByteGAP, a non-continuous graph computing system that leverages PMEM's unique features, such as durability, byte-addressability, memory-like access, lower latency, and high capacity. The non-continuous approach uses checkpointing mechanisms to achieve effective fault detection and recovery. ByteGAP provides two key contributions: (1) lightweight distributed checkpointing based on PMEM, (2) efficient dual-mode PMEM management for optimizing PMEM read and write operations. Moreover, we present a comprehensive evaluation method that demonstrates the system's ability to handle the challenges associated with large-scale computing tasks. The findings lay the foundation for future research in distributed graph computing systems and advocate for a non-continuous approach to graph computing.