Rethinking remote memory placement on large-memory systems with path diversity

Abstract

In this study, we rethink the existing memory placement strategies for multi-chip server systems. Processor vendors such as Intel and AMD have introduced the multi-chip based high density server architecture for large-scale data centers. With the advance of the processor-interconnect, multiple CPU chips are connected through a scalable point-to-point network such as Intel UPI and AMD Infinity Fabric. Nevertheless, the existing operating systems including Linux do not fully take advantage of the processor-interconnect to manage the memory traffic across memory nodes. We present two memory placement techniques exploiting the point-to-point interconnect to improve overall throughput while minimizing the hot-spot problem.