Affinity-on-next-touch: An extension to the linux kernel for NUMA architectures

Abstract

For many years now, NUMA architectures are being used in the design of large shared memory computers and they are gaining importance even for smaller-scale systems. On a NUMA machine, the distribution of data has a significant impact on the performance and scalability of data-intensive programs, because of the difference in access speed between local and remote parts of the memory system. Unfortunately, memory access patterns are often very complex and difficult to predict. Affinity-on-next-touch may be a useful page placement strategy to distribute the data in a suitable manner, but support for it is missing from the current Linux kernel. In this paper, we present an extension to the Linux kernel which implements this strategy and compare it with alternative approaches. © 2010 Springer-Verlag Berlin Heidelberg.