Accelerating large-scale graph neural networks based on solid-state drive through enhancing page cache efficiency

Abstract

As the size of graphs increases, Graph Neural Networks (GNNs) often face memory demands exceeding DRAM capacity. Given this trend, one solution is to utilize SSDs as extended memory while leveraging host DRAM as a cache via the OS page cache. However, the high miss rate and the slow I/O operations of SSDs can hamper performance in such scenarios. We propose the solution, which reduces SSD access frequency by caching frequently used node embeddings in a fine-grained manner on DRAM and exploits the high internal bandwidth of SSDs by offloading a portion of the computations onto SSDs. Our approach shows an average 7.76x performance improvement.