Speeding Up SpMV for Power-Law Graph Analytics by Enhancing Locality & Vectorization Serif Yesil Azin Heidarshenas Adam Morrison Josep Torrellas Dept. of Computer Science Dept. of Computer Science Blavatnik School of Dept. of Computer Science University of Illinois at University of Illinois at Computer Science University of Illinois at Urbana-Champaign Urbana-Champaign Tel Aviv University Urbana-Champaign
[email protected] [email protected] [email protected] [email protected] Abstract—Graph analytics applications often target large-scale data-dependent behavior of some accesses makes them hard web and social networks, which are typically power-law graphs. to predict and optimize for. As a result, SpMV on large power- Graph algorithms can often be recast as generalized Sparse law graphs becomes memory bound. Matrix-Vector multiplication (SpMV) operations, making SpMV optimization important for graph analytics. However, executing To address this challenge, previous work has focused on SpMV on large-scale power-law graphs results in highly irregular increasing SpMV’s Memory-Level Parallelism (MLP) using memory access patterns with poor cache utilization. Worse, we vectorization [9], [10] and/or on improving memory access find that existing SpMV locality and vectorization optimiza- locality by rearranging the order of computation. The main tions are largely ineffective on modern out-of-order (OOO) techniques for improving locality are binning [11], [12], which processors—they are not faster (or only marginally so) than the standard Compressed Sparse Row (CSR) SpMV implementation. translates indirect memory accesses into efficient sequential To improve performance for power-law graphs on modern accesses, and cache blocking [13], which processes the ma- OOO processors, we propose Locality-Aware Vectorization (LAV).