SLA-Guided Energy Savings for Enterprise Servers

Current hardware and operating system power management mechanisms for
servers and data center machines employ control schemes that take little or
no guidance from application software. On these machines, the hardware
and OS take responsibility for controlling frequencies of cores according
to indications of resource utilization. Sophisticated as these schemes are,
they miss the cogent participation of application logic (such as transaction
processing, media, business intelligence, and other services) in determining
precisely how quickly programs must execute to meet service level agreements
(SLAs) or when to defer optional activities. For decades now, server software
designers have innovated and tuned their offerings for best performance,
driven by a competitive landscape that rewarded highest performance with
market success. In practice this has meant software architectures in which
complex and compound purposes are achieved by concurrent progressions
of computations, where external logic cannot slow an isolated resource
without risking unintended consequences to widely mapped computations.
We devised and prototyped one answer to this quandary: a solution we
call an SLA Governor, that engages software in the management of its own
power-performance tradeoffs. In continuous conversation with software,
this solution puts into effect power consumption changes that converge
software performance with that expected by an SLA. Compared to a welltuned
baseline system with a default OS–based power management, the SLA
Governor provides 8.3 percent better energy savings for an online transaction
processing workload and 7.3 percent better power normalized performance
for the SPECpower benchmark. These gains are obtained on two recent
generation machines with Intel E5-5500 (codename Nehalem) and Intel
E5-5600 (codename Westmere) series processors, while executing the same
load and maintaining the application’s SLA.

Download (2.23 MB)