PowerTune: Differentiated Power Allocation in Over-Provisioned Multicore Systems

Exponentially increasing transistor density with each processor generation, along with constant chip-level power budgets and a slower rate of improvement in transistor power dissipation, exponentially decreases the percentage of transistors that can switch on simultaneously. Such ´over-provisioned multicore´ systems require active power management technologies to maintain normal operation constraints. Given the increasing use of shared virtualized infrastructure, these power management capabilities must address the highly diverse behavior of different applications in their processor usage and performance requirements. This paper presents a solution ´Power Tune´ to perform software managed QoS-aware power allocation for over-provisioned multicore processors. Power proportions can be user-specified or dynamically determined based on QoS-policy. When doing so, platform power constraints are modeled using ´power credits´ allocated to virtual machines. Budget constraints are enforced by repeatedly transitioning the CPU into active and idle-states. The solution is implemented in the Xen hypervisor by augmenting its existing credit scheduler. Experimental evaluations are carried out using multiple allocation policies (HWAlloc, FixAlloc, and QoSAlloc), with results showing effectiveness of software-controlled power allocation for VMs with diverse performance requirements and dynamic execution behavior.