SecBCache: Evaluating Banked Interleaved Word-Granularity Sector Cache for gather/scatter Operations on Sparse Data

Abstract Sparse workloads in scientific computing, graph analytics, and machine learning rely on indirect memory accesses (A[B[i]]). These accesses have little spatial locality, so caches use few words of each line they fetch. We present SecBCache, a word-granularity sector cache whose interleaved banks can serve different elements of a vector gather/scatter. It fetches and evicts 8 B subsectors that share one tag per sector. We evaluate SecBCache in gem5 with breadth-first search (BFS) and sparse general matrix–matrix multiplication (SpGEMM) on 16 SuiteSparse datasets. Compared with an 8-way baseline cache, SecBCache reduces SRAM data-array writes by a median of 63% for BFS and 80% for SpGEMM. It also lowers L1 dynamic energy for every dataset. Its area is at most 2.46% of the cache hierarchy. An analytical bank-conflict model conservatively estimates median memory-access delay reductions of 26% for BFS and 50% for SpGEMM. These are modeled delays, not application speedups.