Authentication of Sub-NUMA Clustering effect on Intel Skylake for Memory Latency and Bandwidth

Abstract

The search efficiency of in-memory databases depends significantly on how quickly DRAM data can be obtained. With an increasing multiple core on processors, it is more difficult to satisfy the requirement that all attached DRAM on a processor is equally accessible to any core. Intel has therefore implemented a Sub-NUMA Clustering (SNC) mode on Skylake, which subdivides cores and memories into various sub for enhanced core-to-memory access within each processor sub-domain. Similar modes are given by other models. When an in-memory database shares data between staff on cores in separate sub-domains, the use of SNC creates problems in how to manage database workloads among domains. In this research, we verify the effect of SNC specifically on Intel Skylake on memory latency and bandwidth. We conclude that two similarly broad analytical workloads are focused on different and fully independent sub-domains just up to 3 percent will improve query throughput and be totally isolated from each other. Often, as memory bandwidth is split into sub-domains rather than aggregated across the entire processor, bandwidth-sensitive analytical workloads significantly reduce application performance if data is not split equally among sub-domains.