TR-2018-03
Handing DVFS to Hardware: Using Power Capping to Control Software Performance
Connor Imes; Huazhe Zhang; Kevin Zhao; Henry Hoffmann. 1 February, 2018.
Communicated by Henry Hoffmann.
Abstract
Dynamic voltage and frequency scaling (DVFS) has been the
cornerstone of innumerable software approaches to meeting
application performance requirements with minimal energy.
However, recent trends in technology—e.g., moving voltage
converters on chip—favor hardware control of DVFS,
as hardware can both react faster to external events and perform
fine-grained power management across a device. We
respond to these trends with CoPPer, which instead uses
hardware power capping to meet application performance
requirements with high energy efficiency.We find that meeting
performance requirements with power capping is more
challenging than using DVFS because the relationship between
power and performance is non-linear and has diminishing
returns at high power values. CoPPer overcomes these
difficulties by using adaptive control to approximate nonlinearities
and a novel gain limit to avoid over-allocating
power when it is no longer beneficial. We evaluate CoPPer
with 20 parallel applications on a dual-socket, 32-core
server system and compare it to both a simple linear DVFS
controller and to an existing control-theoretic, model-driven
software DVFS manager. CoPPer does not require a userspecified
model or application pre-characterization, yet provides
all the functionality of the sophisticated DVFS-based
approach. Compared to DVFS, CoPPer’s gain limit reduces
energy by 6% on average and by 12% for memory-bound
applications. For high performance requirements, the energy
savings are even greater: 8% on average and 18% for
memory-bound applications.
Original Document
The original document is available in PDF (uploaded 1 February, 2018 by
Henry Hoffmann).
