DMG delivers 4.9x speedup for graph processing on disaggregated memory

Traditional graph processing systems rely on monolithic servers with a fixed compute-to-memory ratio, leading to significant resource under-utilization in data centers. To address this, researchers have turned to disaggregated memory (DM) architectures, but existing DM-based graph processing systems are highly impractical—they fail to scale beyond a single memory node and a single compute node, and they require compute-side caches that are orders of magnitude larger than conventional DM practices. Now, Yizou Chen, Tsun-Yu Yang, Zhisheng Hu, Baotong Lu, and Ming-Chang Yang have introduced DMG, a system designed from the ground up to make graph processing on DM both scalable and efficient.

DMG introduces several key innovations. It proposes a DM-friendly graph store with optimized retrieval to improve graph access efficiency over the disaggregated network. To reduce the overhead of update propagation between compute and memory nodes, DMG presents an adaptive update coordinator that orchestrates propagation with minimal latency. For load balancing, DMG employs a two-stage workload manager: a coarse-grained initial partitioning followed by fine-grained runtime re-scheduling. Experimental results show that compared to the state-of-the-art DM-based graph processing system, DMG delivers up to 4.9× better performance and scales elastically across both compute and memory resources. Remarkably, it tames compute-side cache demands by up to 18.9×, making it a practical, deployable solution for real-world data centers dealing with ever-growing graph datasets.