Research on Optimization Methods for Hybrid Memory Graph Processing of Memory Access
Wang Zhun · 2025
To address the memory access issues in graph computations, this paper constructs a hybrid memory system using non-volatile random access memory (NVRAM) and DRAM. We investigate the memory access behaviors of graph data stored in the compressed sparse row (CSR) format within NVRAM, revealing significant differences in access frequency, hit rates, and request types among various graph data. Frequent random read/write requests for attribute data and competition for NVRAM buffer resources among different types of graph data limit the memory access performance of graph computations. In response to these challenges, we propose NDGraph, which utilizes the distinct memory access characteristics of different graph data types and the features of various storage mediums. By implementing an address mapping mechanism and classification logic within the memory controller, NDGraph separates memory data flows. It allocates frequently modified attribute data to DRAM while providing efficient data pathways for different types of graph data within NVRAM through bypassing cache and partitioned access design. This approach prevents irregular accesses from vertex data that could disrupt the locality of neighbor data. Experimental results demonstrate that compared to state-of-the-art solutions, NDGraph achieves an average improvement of 1.11 times in bandwidth utilization under multi-core, multi-channel configurations. Furthermore, the overall system performance improves by 1.17 times in single-core mode and 1.20 times in multi-core mode.