On the Energy Inefficiency of MPTCP for Mobile Computing
 2 Citations
 615 Downloads
Abstract
Mobile devices have embraced MultiPath TCP (MPTCP) for leveraging the path diversity. MPTCP is a doubleedged sword since mobile phones are suffering excessively from short battery life span. In order to find energy efficiency of MPTCP, the signal quality and the transferred file size have been taken into account. We formulate the above problem as a Markovian Decision Process (MDP) for symmetric and asymmetric network traffic. Numerical and simulation results surprisingly show that MPTCP is not efficient in any selected scenarios and the proposed scheme can save 56 % of energy compared to the conventional MPTCP.
Keywords
MPTCP Mobile device Energy MDP1 Introduction
Wide coverage oriented cellular base stations and WiFi hotspots coverage go hand in hand with arming mobile phones by different transmission interfaces such as cellular and WiFi, enhance the users’ ubiquitous connectivity. Due to the errorprone nature of wireless channels, poor wireless signal quality is frequently experienced spatially and temporally by end users. It has been shown that the signal quality in WiFi or cellularbased networks dips in most cases. For instance, 43 % and 21 % of mobile phone data traffic is transmitted during low cellular and WiFi signal strength, respectively [2]. In addition, low signal strength could drain the mobile device battery. Due to the limited nature of mobile phone batteries, meeting high quality of service (QoS), while maintaining reasonable level of energy consumption, is of great importance for endusers. It is noteworthy that reaching high QoS and energy efficiency is not possible, in many cases, due to their conflicting requirements. Thus, designing a proper scheme to obtain reasonable levels of QoS and energy efficiency is critical.

We make a tight relation between signal quality, file size and energy consumption; besides the optimization problem is formulated as a Markovian decision making problem by employing buffer size dynamics to better understand transferring file sizes.

We endeavor to minimize energy consumption by exploiting diverse characteristics of interfaces, and for the first time we have found that MPTCP is not energy efficient in real scenarios, particularly in simultaneous uploading and downloading.

We investigate the effect of energy consumption on asymmetric and symmetric data traffic for a mobile user facing variety of signal strength on different interfaces.
The paper is organized as follows. Section 2 introduces the energy model, and shed light on the effect of file size on MPTCP throughput. Section 3 describes the problem formulation. Section 4 outlines the simulation results, and Sect. 5 reviews the related works. Finally, we conclude in Sect. 6.
2 Proposed Model
The proposed model aims to optimize the transmission energy consumption as a utility function based on signal strength and transferring file size. In the following, we describe the underlying network model and assumptions for the problem at hand.
2.1 Energy Model
Packet transfer coefficients [4]
LTE  WiFi  

Download  \( \alpha ^d \)  10.04  4.64 
\( \beta ^d \)  −0.89  −0.81  
Upload  \( \alpha ^u\)  13.34  3.61 
\(\beta ^u\)  −0.83  −0.66 
2.2 Effect of File Size on Throughput
Upload coefficients
\(\kappa ^u \)  \( \eta ^u \)  \( \iota ^u \)  

TCP over WiFi  −23.35  −0.16  29.48 
TCP over LTE  −27.38  −0.08  29.62 
MPTCP over WiFi  −27.58  −0.14  32.57 
MPTCP over LTE  205.7  0.0052  −203.3 
Download coefficients
\( \kappa ^d \)  \(\eta ^d \)  \(\iota ^d \)  

TCP over WiFi  −24.1  −0.17  30 
TCP over LTE  541  0.0055  −539 
MPTCP over WiFi  −95  −0.039  98.22 
MPTCP over LTE  −14.05  −0.15  18.59 
3 Problem Formulation
3.1 Definitions
A mobile device decision making is based on the assumption that the system dynamics are determined by an MDP. MDP consists of the following elements: decision epochs, states, states transition probabilities, actions and costs. The mobile device can make a decision about tuning one or both interfaces from idle to active mode in each time epoch, where \(\mathcal T =\{1, 2, ..,y\} \) represent a set of decision epoch. The signal quality states are introduced by the following: firstly, a set of signal quality states for interface i, \( \mathcal {R}_i=\{r^{1}_i, r^{2}_i,...,r^{n}_i \}\), where \( r^{1}_i \) is minimum and \( r^{n}_i \) is maximum signal quality state. It is worth adding here that each signal quality corresponds to specific uplink and downlink bandwidth. Secondly, signal quality states transition probability matrix for interface i, \( P^{r_{i}}=[p^{r_{i}}(r^{k}_{i}r^{j}_{i}), r^{1}_{i}\le r^{j}_{i}, r_{i}^k\le r_{i}^n] \). In [11] we have shown that how to find the most probable signal quality states by getting RSSI as an observation and in this paper we assume the signal quality state is known. Let \(\mathcal S = \mathcal R^{1} \times \mathcal R^{2}=\{(r_1^1, r_2^1), (r_1^1, r_2^2), ..., (r_1^n, r_2^n) \}=\{s_1, s_2,..., s_N\}\) represent the total state space, where \( \mathcal N=\mathcal R_{1}\times \mathcal R_{2}\). The action space \(\mathcal A=\{(on, on),(on, off), (off, on)\} \) denotes a set of all possible actions for both interfaces. Without loss of generality, let LTE be the first interface and WiFi be the second; consequently, (on, off) means LTE interface is active and the WiFi interface is idle.
Before discussing decision rules and policies, we have to shed light on cost calculation. In order to get the benefit of high accuracy of throughput estimation we need to have information about file size. Therefore, we use the backlog queue to have a better understanding of transferring file size.
4 Simulation and Results
In order to appraise the proposed algorithm, we have performed the simulations by ns3 (https://www.nsnam.org). The simulation parameters are included in Table 4. We have considered three LTE and three WiFi signal quality states. Based on the selected states, we found nine spots on the simulation field as it noted in Fig. 4. We calculated nine upload and download tables as in Tables 2 and 3. Before the simulation the best policy in each state for different arrival rates have been calculated. After that the traffic has been generated between two nodes, and the mobile node (UE1) moves from one spot to another by using the most efficient policy in each state.
Simulation parameters
LTE RSSI states  (−45, −75), (−76, −105), (−106, −135) 
WiFi RSSI states  (−50, −61), (−62, −75), (−76, −90) 
Point to Point bandwidth  100 Mbps 
Point to Pint delay  1 ms 
Direct Code Execution (DCE)  version 1.7 
From Figs. 5, 6, 7 and 8, the optimal policies have been marked by a red arrow. Taking the advantage of the packet capture feature of NS3, we have plotted the energy consumption behavior of other policies in each state. SAMPTCP optimal decision was completely compatible with real simulation and surprisingly MPTCP has not been chosen in any scenario. In fact, this is the first time that uploading and downloading have been considered together, and the results contrast prior research such as [4, 6]. Most of the time either WiFi or LTE are the best policy and even according to Fig. 6 WiFi is the best option for uploading small file size and downloading large file size. SAMPTCP energy consumption outperforms other platforms.
Comparing SAMPTCP to MPTCP, WiFi, and LTE, SAMPTCP saves \( 56\,\%\), \( 16\,\% \) and \( 65\,\%\) energy, respectively. On the other hand, MPTCP is 23 % more throughput efficient than SAMPTCP. However, SAMPTCP achieves a higher throughput compared to WiFi and LTE (Fig. 9).
5 Related Works
MPTCP has been investigated widely from different perspective; nonetheless, energy efficiency attracts the growing interest of researchers. Comparison of TCP over LTE and WiFi with MPTCP in terms of energy efficiency have been discussed in [6] by modeling, measuring and simulating. The authors proposed eMPTCP, which manages subflows expected energy efficiency based on available throughput. In fact eMPTCP is the most similar work to us, but the difference is eMPTCP only considered the downloading case and doesn’t take the uploading energy into account. In addition, to estimate the available throughput eMPTCP sampled downloaded bytes, which introduces the overhead to a network.
GreenBag [1] was proposed as a bandwidth aggregate middleware on wireless links. It estimates the available bandwidth on wireless links and specifies the amount of traffic allocation while taking QoS into account. GreenBag reduces energy consumption between 14 % to 25 %; however, it needs major applications modification which is not feasible. Peng et al. [6] designed a MPTCP algorithm, decreasing energy consumption up to 22 % by intelligently compromising between throughput and energy. However, they only considered bulk data applications such as video streaming and file transfer, and they did not consider the effect of signal strength on throughput.
Markov decision process has been applied by [7] to utilize the optimal interface for different applications. Nevertheless, authors only considered throughput and stationary applications which is infeasible in dynamic environment. For instance, each application has different traffic patterns, and modeling all applications traffic is not practical. Furthermore, the optimal throughput performance is only achieved by one path, and more importantly the system performance would dwindle immensely when the optimal path is the most congested one. Energy aware MPTCP is proposed in [8] to balance the increasing throughput with the energy consumption. Unfortunately, data has been offloaded simply to WiFi whenever possible since they assumed WiFi consumes less energy than LTE. That assumption may not be true in many scenarios. In fact, when a mobile device experiences a weak WiFi signal, not only a huge amount of packet loss and delay would be irrefutable [5], but also as we showed, energy cost would increase sharply.
Lim et al. [5] proposed an inspiring algorithm, MPTCPMA, which manages the path establishment by taking cross layer information, such as Mac layer behavior. They endeavored to increase the WiFi throughput efficiency up to 70 %; nonetheless, energy efficiency hasn’t been taken into account, and they haven’t addressed cellular networks behavior. In contrast to most studies, Tailless MPTCP (TMPTCP) [10] showed that MPTCP energy efficiency could be enhanced by the tail energy minimization. Since the tail energy has a enormous impression on the energy contribution, TMPTCP optimizes jointly energy and delay. However, TMPTCP deals with static bandwidth and the offline experiment demands more investigation.
6 Conclusion
In this paper, we studied the problem of energy efficiency in the context of multipath TCP. We proposed a control mechanism based on Markov Decision Process to distribute traffic over diverse paths while optimizing energy. Performance of the proposed control mechanism was evaluated and compared with the different platforms. The results indicate that the proposed control mechanism significantly outperforms other schemes in all considered scenarios and MPTCP is not energy efficient in any selected cases. Even though we have applied our approach to LTE, the model could simply extend to 3G networks. The effect of delay would be investigated in future work.
References
 1.Bui, D.H., Lee, K., Oh, S., Shin, I., Shin, H., Woo, H., Ban, D.: Greenbag: energyefficient bandwidth aggregation for realtime streaming in heterogeneous mobile wireless networks. In: IEEE 34th RealTime Systems Symposium (RTSS 2013), pp. 57–67. IEEE (2013)Google Scholar
 2.Ding, N., Wagner, D., Chen, X., Pathak, A., Hu, Y.C., Rice, A.: Characterizing and modeling the impact of wireless signal strength on smartphone battery drain. ACM SIGMETRICS Perform. Eval. Rev. 41, 29–40 (2013). ACMCrossRefGoogle Scholar
 3.Ford, A., Raiciu, C., Handley, M., Barre, S., Iyengar, J., et al.: Architectural guidelines for multipath TCP development. IETF, Informational RFC 6182, 20701721 (2011)Google Scholar
 4.Lim, Y.S., Chen, Y.C., Nahum, E.M., Towsley, D., Gibbens, R.J.: Improving energy efficiency of MPTCP for mobile devices. arXiv preprint arXiv:1406.4463 (2014)
 5.Lim, Y.S., Chen, Y.C., Nahum, E.M., Towsley, D., Lee, K.W.: Crosslayer path management in multipath transport protocol for mobile devices. In: Proceedings of INFOCOM, pp. 1815–1823. IEEE (2014)Google Scholar
 6.Peng, Q., Chen, M., Walid, A., Low, S.: Energy efficient multipath TCP for mobile devices. In: Proceedings of the 15th ACM International Symposium on Mobile Ad hoc Networking and Computing, pp. 257–266. ACM (2014)Google Scholar
 7.Powell, W.B.: Approximate Dynamic Programming: Solving the Curses of Dimensionality, vol. 703. Wiley, New York (2007)CrossRefzbMATHGoogle Scholar
 8.Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley, New York (2014)zbMATHGoogle Scholar
 9.Schulman, A., Navda, V., Ramjee, R., Spring, N., Deshpande, P., Grunewald, C., Jain, K., Padmanabhan, V.N.: Bartendr: a practical approach to energyaware cellular data scheduling. In: Proceedings of the Sixteenth Annual International Conference on Mobile Computing and Networking, pp. 85–96. ACM (2010)Google Scholar
 10.Shamani, M.J., Zhu, W., Naghshin, V.: TMPTCP: Tailless Multipath TCP. In: 10th International Conference on Broadband and Wireless Computing, Communication and Applications (BWCCA), pp. 325–332 (2015). doi: 10.1109/BWCCA.2015.103
 11.Shamani, M.J., Zhu, W., Rezaie, S., Naghshin, V.: Signal aware multipath TCP. In: 2016 Proceedings IEEE/IFIP of WONS, pp. 104–107. IEEE/IFIP (2016)Google Scholar
 12.Wischik, D., Raiciu, C., Greenhalgh, A., Handley, M.: Design, implementation and evaluation of congestion control for multipath TCP. In: NSDI, vol. 11, p. 8 (2011)Google Scholar