Effective Task Scheduling of In-Memory Databases on a Sub-NUMA Processor Topology

Sammanfattning: Query throughput of in-memory databases greatly depends on how fast data can be accessed in DRAM. With a growing number of cores on processors, the expectation that all attached DRAM on a processor is equally accessible by every core is harder to meet. Therefore, Intel has introduced a mode coined SubNUMA Clustering (SNC) on Skylake, which subdivides cores and memories into multiple sub-domains for improved core-to-memory access within each sub-domain on the processor. Other models offer similar modes. Usage of SNC poses challenges on how to balance database workloads between domains when an in-memory database share data between workers on cores in different sub-domains. In this thesis, we verify SNC’s impact on memory latency and bandwidth primarily on Intel Skylake. Furthermore, we test the in-memory database SAP HANA when using SNC on single-row and analytical workloads. We conclude that two equally sized analytical workloads put on separate sub-domains and fully isolated from each other, only would increase query throughput up to 3%. Also, as memory bandwidth is divided into sub-domains rather than aggregated on the entire processor, analytical workloads that are sensitive to bandwidth drastically decrease query throughput if data are not evenly partitioned between sub-domains.