An Operating System for Clumpy Shared Memory

Abstract

Despite the pervasive expansion of parallel computer architectures and the expected exponential growth of cores on a single chip into the thousands, it is difficult to efficiently scale the performance and storage overhead of processor caches using conventional cache coherence techniques. One solution to this dilemma of scalable cache coherence is to restrict the domain of coherence to a subset of global memory. A novel technique, known as Clumpy Shared Memory (CSM), attempts to do just that by restricting either the total number of sharers that can access a coherence domain or the total number of home nodes that can do so. This work seeks to develop the software infrastructure necessary for CSM by building the operating systems support and interface for the underlying hardware implementation. We use the open-source and widely popular Linux operating system as a starting point for this OS, and provide a detailed discussion of previous and related work, design choices, methodology, and implementation.