Energy Proportional Datacenters

Daniel Wong
Ph.D. Candidate
Electrical Engineering, University of Southern California, Los Angeles
SERC 306
Friday, February 27, 2015 - 11:00
Datacenters provide the infrastructure backbone necessary to support big data analytics and cloud services, which are increasingly employed to tackle a diverse set of grand challenges. But datacenter power consumption is growing at an unsustainable pace. In order to keep up with the hyperscale growth in datacenter demand, it is imperative that datacenters become more energy efficient. Servers, the largest power consumer in datacenters, are optimized for high energy efficiency only at peak and idle load, but rarely operate in that region. Therefore, there is a need for energy proportional computing, where servers consume power in proportion to their utilization. How to achieve or surpass ideal energy proportionality is the focus of this talk.     
Toward this goal, I will first present a historical trend analysis of energy proportionality, using novel metrics, in order to identify opportunities for proportionality improvements. Second, I will present KnightShift, a heterogeneous server architecture that tightly couples a low-power Knight node with a high-power primary server, which achieves near-ideal energy proportionality. Finally, I will present the implications of high energy proportional servers on cluster-level energy proportionality. We find that traditional cluster-level energy proportionality techniques may actually limit cluster-wide energy proportionality, and it may now be more beneficial to depend solely on server-level low power techniques such as KnightShift; a finding that is a major departure from conventional wisdom.