Experimental evaluation of NUMA effects on database management systems

Abstract

NUMA systems with multiple CPUs and large main memories are common today. Consequently, database management systems (DBMSs) in data centers are deployed on NUMA systems. They serve a wide range of database use-cases, single large applications having high performance needs as well as many small applications that are consolidated on one machine to save resources and increase utilization. Database servers often show a natural partitioning in the data that is accessed, e.g., caused by multiple applications accessing only their data. Knowledge about these partitions can be used to allocate a database's memory on the different nodes accordingly: a strategy that increases memory locality and reduces expensive communication between CPUs. In this work, we show that partitioning a database's memory with respect to the data's access patterns can improve the query performance by as much as 75%. The allocation strategy is enabled by knowledge that is available only inside the DBMS. Additionally, we show that grouping database worker threads on CPUs, based on their data partitions, improves cache behavior, which in turn improves query performance. We use a self-developed synthetic, low-level benchmark as well as a real database benchmark executed on the MySQL DBMS to verify our hypotheses. We also give an outlook on how our findings can be used to improve future DBMS performance on NUMA systems.