Performance evaluation of main-memory hash joins on KNL

New hardware features have propelled designs and analysis in main-memory hash joins. In previous studies, memory access has always been the primary bottleneck for hash join algorithms. However, there are relatively few studies devoted to bottlenecks analysis on knights landing processor (KNL). In this paper, we pay attention to the state-of-the-art hash join algorithms on KNL and analyse their bottlenecks under different workloads. The analysis and comparisons in the paper show that both memory latency and bandwidth are keys to improve hash joins, and multi-channel dynamic random access memory (MCDRAM) reasonably plays a vital role in enhancing performance. Notably, we find that hash join algorithms that are hardware-oblivious perform better than hardware-conscious approaches. A typical algorithm of hardware-oblivious joins achieves a better performance than ever before to the best of our knowledge. Through the analysis, we shed light on how new features of KNL affect the performance of hash joins.