Coordinated Optimization: Dynamic Energy Allocation in Enterprise Workload

Controlling how much power server machines draw has become increasingly
important in recent years. The accuracy and agility of three types of actions
are critical in power governance: (1) selecting which hardware elements must
run at what rates to meet performance needs of software, (2) assessing how
much power must be expended to achieve those rates, and (3) adjusting the
power outlay in response to shifts in computing demand. Observing how
variations in a workload affect the power drawn by different server components
provides data critical for analysis and for building models relating quality of
service expectations to power consumption. This article describes a process of
observation, modeling, and course corrections that is successful in achieving
autonomic power control in an Intel® Xeon®E5-2600 server machine meeting
varying response time and throughput demands during the execution of a
database query workload. The process we describe in the article starts with
fine-grained power-performance observations permitted by a distributed set
of physical and logical sensors in the system. These observations are used to
train models for various phases of the workload, with accuracy between 97 and
98.5 percent. Once trained, system power, throughput, and latency models
participate in optimization heuristics that redistribute the power to maximize
the overall performance/watt of the server.

Download (764.86 KB)