PI-RATE: QoS and Performance Management in CMP Architectures

摘要

As new multi-threaded usage models such as virtualization and consolidation take advantage of multiple cores in CMP architectures, the impact of shared resource contention between VMs and user-level applications introduces Quality of Service(QoS) concerns and challenges. QoS-aware management of these shared platform resources is therefore becoming increasingly important. Various QoS schemes for resource management have been recently proposed, but most of these prior efforts have been focused on controlling individual resource allocation based on priority information passed down from the OS or Hypervisor to system resources. The complexity of this approach increases when multiple levels of resources are associated with an application's performance and power consumption. In this paper we employ simpler rate-based QoS mechanisms which control the execution rate of competing applications. To enable differentiation between simultaneously running applications' performance and power consumption, these rate mechanisms need to dynamically adjust the execution of application. Our proposed PI-RATE architecture introduces a control-theoretic approach to dynamically adjust the execution rate of each application based on the QoS target and monitored resource utilization. We evaluate three modes of PI-RATE architecture - cache QoS targets, performance QoS targets and power QoS targets - to show that the PI-RATE architecture is flexible and effective at enabling QoS in a CMP platform.