Linux Support for Memory Traffic Shaping

Abstract

Although parallel computer architectures have become ubiquitous, mem- ory bandwidth places severe limits on the scalability and performance of these systems as the number of cores on a single chip grows. A potential way to address this critical issue has been designed by the Princeton Parallel Group { MITTS (Memory Inter-arrival Time Tra c Shaping), a distributed hardware mechanism that classifies and shapes tra c between each core and main memory based on the time between successive requests. Shaping mem- ory tra c on a per-core basis enables fine-tuned bandwidth allocation and increases both efficiency and fairness for multi-program workloads. This work develops the operating system-level software necessary to support and fully exploit the capabilities that MITTS provides and to test its performance un- der datacenter-like workloads. We use the Princeton Parallel Group's 25-core Piton processor as a hardware platform and the Linux operating system as the starting point for our development. From there, we successfully imple- ment and test a comprehensive Linux subsystem that allows users to securely configure MITTS on a per-thread or per-user basis with a simple, familiar system call interface.