Energy Efficiency Analysis of Heterogeneous
Platforms: Early Experiences
Youhuizi Li
∗†
, Weisong Shi
†
, Congfeng Jiang
∗
, Jilin Zhang
∗
and Jian Wan
∗‡
∗
Key Laboratory of Complex Systems Modeling and Simulation, Hangzhou Dianzi University, China
†
Mobile and Internet Systems Laboratory, Wayne State University , USA
‡
School of Information and Electronic engineering, Zhejiang University of Science & Technology, China
{huizi,cjiang,jilin.zhang,wanjian}@hdu.edu.cn, weisong@wayne.edu
Abstract—Heterogeneous multi-core platforms, e.g., ARM’s
big.LITTLE, are a promising trend to improve the performance
and energy efficiency of future mobile systems. However, the
immediate benefits and the challenges to take advantage of the
heterogeneity are still not clear. In this paper, we present our ear-
ly experiences about the energy efficiency of the two big.LITTLE
heterogeneous platforms: ODROID XU+E and ODROID XU3.
We quantified compared them with homogeneous platforms
through multiple benchmarks which include popular mobile
applications and high-performance parallel benchmarks. Besides,
we analyzed the scheduling impact on the energy consumption
of the heterogeneous platforms and the migration cost is also
discussed. Based on the results, several insights, such as fine-
granularity power control and thread level parallelism, related
to hardware, application and system design are derived.
Index Terms—Heterogeneous platform, Energy consumption,
big.LITTLE
I. INTRODUCTION
As the result of the dark silicon issue and increasing
demand of specialized components, heterogeneity becomes
more and more important and leads the trend of future devices’
development, especially for mobile platforms. Heterogeneity
is a general concept that may refer to CPU/GPU computing
architecture, mixed types of accelerators and so on. The
advantage of heterogeneous platforms is that they can improve
energy efficiency while maintaining performance[1], [2]. To
evaluate their benefits, previous work usually leverages DVFS
to simulate different types of CPUs [3]. With the emerging
of the ARM big.LITTLE processor [4] and Samsung Exynos
5 Octa system-on-chip (SoC) [5], we have the opportunity to
explore and exploit real heterogeneous hardware platforms.
In this paper, we undertake the following questions:(1)
Compared with homogeneous platforms, how much energy can
be saved in heterogeneous platforms? We want to know the
capability of heterogeneous platforms. (2) what is the impact
of scheduling from energy aspect? As the different types of
processor exist, the scheduling algorithm takes the responsi-
bility of choosing proper processor for different workloads.
Both of the benefits can be gained from the correct scheduling
and penalty of the improper scheduling are important. And
finally (3) what is the migration overhead on performance
and energy? Applications have different phrases (e.g. loading
content, waiting user input, etc.) and scheduler needs dy-
namically migrate workload to proper cores in runtime. In
this situation, migration overhead is one of the key factors
to decide when and which core to migrate. All of the three
questions directly influence the benefits we can get from
heterogeneous platforms.
To answer these questions, we picked the two ARM
big.LITTLE platforms, ODROID XU+E (XUE) and ODROID
XU3 (XU3) [6], as experimental platforms to study. Both
XUE and XU3 have heterogeneous cores but provide different
control strategies. XUE offers cluster switching and XU3
supports heterogeneous multi-processing [5]. The system and
component level power information are collected to analyze
their energy behavior under various cases. By analyzing the
results, we can further understand the characteristics of the
big.LITTLE platforms and wisely use them.
In this paper, we have three main contributions:
• We investigate the two generations of the same hetero-
geneous platforms and homogeneous platforms from the
viewpoint of energy in a quantitative way. We found that
most of the energy savings in heterogeneous platforms
come from idle time and there is no much benefit for
sustained heavy workloads.
• We analyze the impact of scheduling and migration
overhead on the ARM big.LITTLE devices from the per-
formance and energy aspects. The improper scheduling
can consume 5% - 30% more energy.
• We derived a list of insights, such as fine-granularity
power control and thread level parallelism, related to
hardware, application and operating system design.
The remainder of the paper is organized as follows: We
introduce the two heterogeneous platforms and illustrate our
experiment setup in Section II. The detailed case studies are
demonstrated in Section III, which compared the heteroge-
neous and homogeneous platforms and analyzed the impact
of scheduling and migration cost. Following that, we discuss
a list of insights in Section IV. The related work is presented
in Section V and Section VI summarizes the paper.
II. EXPERIMENT SETUP
To practically investigate heterogeneous platforms, we did
experiments on two generations of ARM big.LITTLE plat-
forms produced by Hardkernel [6]. Their specifications are
presented in Table I. ARM’s big.LITTLE processor [4] is a
single-ISA heterogeneous multi-core processor which contains
978-1-5090-5117-5/16/$31.00
c
2016 IEEE