TR-2016-10
Big Data for LITTLE Cores: Combining Learning and Control for Mobile Energy Efficiency
Nikita Mishra; Connor Imes; Huazhe Zhang; John D Lafferty; Henry Hoffmann. 20 September, 2016.
Communicated by Henry Hoffmann.
Abstract
Mobile systems must deliver performance to interactive
applications while simultaneously conserving resources
to extend battery life. There are two central challenges to
meeting these conflicting goals: (1) the complicated optimization
spaces arising from hardware heterogeneity and
(2) dynamic changes in application behavior and resource
availability. Machine learning techniques handle complicated
optimization spaces, but do not incorporate models
of system dynamics; control theory provides formal guarantees
of dynamic behavior, but struggles with non-linear
system models. In this paper, we propose CALOREE, a
combination of learning and control techniques to meet
performance requirements on heterogeneous devices in
unpredictable environments. CALOREE combines a hierarchical
Bayesian model (HBM) with a lightweight control
system (LCS). The HBM runs remotely, learning customized
performance/power models. The LCS runs on the
mobile system and tunes resource usage to meet performance
goals. The Performance Hash Table (PHT) is the
interface between the two and allows the LCS to apply
the learned models in constant time. We test CALOREE’s
ability to manage ARM big.LITTLE systems. Compared to
existing learning and control methods, CALOREE delivers
more reliable performance – only 2% error compared
to 4.5-5.4% for learning and 4.7% for control – and lower
energy – within 7% of optimal on average as compared to
25-52% for learning and 26% for control. Furthermore,
we demonstrate CALOREE’s ability to meet performance
and energy goals in dynamic systems with phase changes
and multiple applications running on the same system.
Original Document
The original document is available in PDF (uploaded 20 September, 2016 by
Henry Hoffmann).
