active wireless network switching system is proposed in [40] to automatically select the best wireless connection for data
transmission on the smartphone. A recent work, Adapp [6], is capable to optimally select the network service that suits an
application best in terms of user-desired quality of experience (QoE). However, these methods focus predominantly on
throughput enhancement rather than energy saving.
Current mobile devices are severely constrained by limited battery capacity. It has long been revealed that the power
drained by network interfaces constitutes a large fraction of the total power used by a mobile device [36]. Such a situation
becomes even worse with the rising popularity of mobile cloud computing [18]. Hence, recent efforts have been made on
transmission scheduling of radio interfaces to improve energy efficiency of mobile devices. Some of them are proposed
for mobile devices with a single wireless interface. Neely [25] developed a joint strategy (EECA) for power allocation and
admission control that satisfies the constraints on transmit power while maximizing the throughput metric. Based on empir-
ical studies on tail energy (i.e., energy consumed in high-power state after data transfer is completed), TailEnder [5] aims to
aggregate small transmissions into large ones through prefetching and delayed transfer, so that the occurrence of tails (and
thus energy consumption) can be reduced. Similar schemes include TailTheft [19] and Traffic-Backfilling [15]. On the other
hand, some prior studies try to achieve energy efficiency by scheduling data transfer among several available interfaces on
mobile devices. CoolSpots [30] aims to reduce the energy consumption by intelligently deciding whether and when to use
WiFi and Bluetooth based on an application’s bandwidth requirement. Context-for-Wireless [33] employs the statistical
information of historical context to decide whether and when to power-on the WiFi interface to improve the energy effi-
ciency of data transfer in cellular networks. Both SALSA [32] and eTime [35] apply Lyapunov optimization techniques to
make online transmission decisions to balance the tradeoff between energy consumption and transmission delay. However,
these four schemes are interested in determining the lowest energy link among a set of available links at a given instant,
without taking any application requirement on network throughput into consideration.
2.2. Lyapunov optimization techniques for stochastic systems
Lyapunov optimization [27] is a recently developed technique for solving problems of joint system stability and perfor-
mance optimization on stochastic networks, especially communication and queueing systems. To achieve this goal, network
algorithms are designed to make control actions that greedily minimize a bound on the following drift-plus-penalty expres-
sion in each time slot t:
D
ðtÞþVcðtÞ
where
D
ðtÞ (Lyapunov drift) represents the congestion state of queue backlog, cðtÞ denotes the objective function to be opti-
mized, and V is a non-negative weight that is chosen as desired to affect a performance tradeoff between backlog reduction
and penalty minimization. Unlike Markov Decision Process [38] and Dynamic Programming [1], Lyapunov optimization does
not require knowledge of the statistics of related stochastic models, but instead the queue backlog information, to make
online control decisions. These two traditional techniques suffer from the so-called ‘‘curse of dimensionality’’ problem
[27], and result in hard-to-implement systems where significant re-computation might be needed when statistics change
[37]. In contrast, Lyapunov optimization algorithms commonly have a better computational complexity, and are easy to
be implemented in practical systems [32,35]. By now, this new technique has been applied in solving many stochastic net-
work optimization problems, including workload/resource scheduling among data centers [39,31,43], power management in
smart grid [37,10,41], and energy/throughput optimization for wireless systems [26,32,35]. Among them, the work most
relevant to ours is that in [43], which makes online decisions on request admission control, routing and virtual machine
scheduling to balance the tradeoff between the throughput performance and power consumption in SaaS (i.e., Software-
as-a-Service) clouds. This work inspires us to exploit a similar throughput-energy tradeoff issue in mobile cloud computing
scenarios.
3. Basic throughput-energy tradeoff model
As shown in Fig. 1, we consider a mobile device user who has M heterogeneous applications m 2f0; 1; 2; ...; Mg
running on a cloud platform [34]. The whole system operates in discrete time with unit time slots t 2f0; 1; 2; ...g. The data
Fig. 1. The system model for MOTET.
W. Fang et al. / Information Sciences 283 (2014) 79–93
81