DYHR: An Efficient Cache Replacement Strategy for Gnn Based on Hybrid Memory

Graph Neural Network (GNN) runs with a large number of random accesses, resulting in a low cache hit rate, and requires frequent disk accesses when running larger graph datasets, which puts a lot of pressure on the storage system. Hybrid memory architecture by DRAM and NVM (Non-volatile Memory) solves the problem of needing frequent disk accesses with largescale graph datasets. However, the low cache hit rate in the hybrid architecture still makes the GNN run inefficiently. If a reasonable cache replacement strategy can be designed for GNN algorithms in the hybrid architecture to improve the data locality, the cache hit rate and the overall performance of GNN will be greatly improved. In this study, we propose Dynamic Hierarchical Weighted Cache Replacement (DyHR) strategy. DyHR optimizes the distribution of hot and cold data by combining the node degree, access frequency, and community characteristics, designing a hierarchical cache weight model, and improving the multilevel cache hit rate and reducing the average storage access time. The experimental results show that compared with the traditional cache replacement strategies, DyHR improves the hit rate and data access latency of all levels of caching on Cora, PubMed and other datasets to different degrees.