ECG: Expressing Locality and Prefetching for Optimal Caching in Graph Structures

Despite state-of-the-art caching strategies, graph analytics pose a significant challenge for prefetching and replacement policies, as their access patterns are often random with low locality. Prior attempts at addressing these issues involved architectural solutions such as specialized graph prefetchers, compressed graph representation stored in a reserved cache section, or graph reordering for custom replacement policies. However, combining these solutions required extensive modifications to the processor architecture. In this paper, we propose Expressive Caching for Graphs (ECG), which masks the graph clustering knowledge with previous cache optimization techniques into the unused bits (ECG bits) of the graph vertex ID. The processor pipeline's specialized graph addressing mode can extract the vertex ID from the graph edge list to request property data while utilizing the masked ECG bits to determine the replacement policy for better caching and prefetching the next related vertex.