IST Lunch Bunch

Data centers are at the heart of today's computing. Data center servers are typically over-provisioned, due to the quality of service and availability concerns. As a result servers operate at low-mid utilization ranges. Due to the lack of energy proportionality, server power efficiency (instructions executed per Watt) is poor at low utilization. In this talk I will first present simple metrics to quantify energy proportionality that enable system designers to analyze and understand the energy consumption of a server at various utilization levels. Using these metrics we quantify how server energy proportionality progressed over the last decade. This analysis shows that energy proportionality improvements are not uniform across various server utilization levels. In particular, the energy proportionality of even a highly proportional server suffers significantly at low utilization. 

In the second part of this talk we tackle the lack of energy proportionality at low utilization by using server-level heterogeneity to provide an active low power mode. We present KnightShift, where a high-performance server is tightly-coupled with a low-power compute node called the Knight node. Knight responds to low utilization requests while the primary server responds only to high utilization requests, thereby enabling two energy-efficient operating regions. KnightShift is evaluated against a variety of real-world datacenter workloads using a combination of prototyping and simulation, showing up to 75% energy savings with tail latency bounded by the latency of the Knight.