Adaptive subflow allocation for multipath data transmission in mobile edge networks
Introduction
Nowadays most smart phones and laptops are equipped with multiple network interfaces including WiFi, 3G/4G, Bluetooth, etc. However, even multiple network connections are available, operating systems typically use one interface at a time for data transmission and leave the rest interfaces idle. To make full use of the capacity of multiple interfaces, multi-path transmission techniques have been proposed to transfer data over several network connections simultaneously, which attempt to aggregate the resource of multiple heterogeneous access networks [1], [2], [3], [4], [5]. The Multipath TCP (MPTCP) protocol [5], [6] has recently been standardized by the IETF as an ideal solution for multi-path data transmission. MPTCP extends conventional TCP by establishing several subflows over multiple interfaces for a single application, which is shown to be compatible with existing networks [6], and achieves robustness and good performance in data center networks [7] and heterogeneous wireless networks [8], [9], [10], [11].
Numerous studies have focused on theoretical analysis and experimental evaluation of the performance of MPTCP for long-lived flows. Congestion control mechanisms were studied in [12], [13], [14], where a number of congestion control algorithms such as EWTCP, Coupled [12], OLIA [13], and Balia [14] were proposed to achieve better fairness and better throughput by adjusting the congestion window size considering different network characteristics. Subflow scheduling issues were discussed in [15], [16], and several works tried to develop schedulers such as Round-Robin (RR) [16], BLEST [17], ECF [18], DEMS [19], STMS [11], ReLes [20], etc, to split and distribute data flow to the subflows to improve data rate and delay of MPTCP. However, management of the number of subflows has not drawn much attention in the past. Current implementation of MPTCP provided static subflow management strategies [21] which allows user to establish one subflow on each interface (the “fullmesh” mode) or to establish subflows on the default interface (the “ndiffports” mode) by a predefined system configuration. The user is not allowed to change the number of subflows without resuming MPTCP system service. In a word, none of the literature has addressed dynamic subflow management to adjust the number of subflows on each network interface in accordance with the change of network conditions self-adaptively.
In this paper, we illustrate that static subflow management strategy yields sub-optimal MPTCP throughput, and a large performance gain can be achieved by allocating different number of subflows to different network interfaces according to network characteristics. We propose theoretical analysis on the performance of MPTCP under the situation of ideal communication links, and extend the results to the wireless environment with random packet dropping. Based on the analysis, we propose a metric called throughput gain to represent the performance improvement of using multiple subflows. We then present an adaptive subflow allocation algorithm, which periodically measures network conditions and optimizes the throughput gain by adaptively determining the number of subflows for each interface. We implement the proposed algorithm in the Linux Kernel as a new “adaptive” mode of path-manager for MPTCP. We conduct extensive experiments on a real-deployed wireless network testbed, which show that the adaptive strategy greatly improves the throughput of the current MPTCP implementation.
The contributions of this paper are summarized as follows.
- •
We identify the issues of adaptive subflow management in MPTCP, and illustrate that the number of subflows established on each network interface has significant influence on the overall performance of multi-path transmission.
- •
We present theoretical analysis on the average throughput of establishing multiple uncoupled subflows on each interface for long-lived MPTCP connections in wireless networks with lossy links. Specifically, our analysis shows that when the number of subflows is small and the network packet loss rate is high, establishing more sunflows can significantly improve the throughput of MPTCP against performance degradation in lossy network environment.
- •
We propose an adaptive subflow allocation algorithm to determine the proper number of subflows for each network interface. The proposed algorithm adopts a measurement-based dynamic subflow management mechanism to optimize the throughput gain of mobile clients.
- •
We implement the proposed algorithm, integrate it into the MPTCP Linux Kernel implementation, and evaluate its performance on a wireless network testbed. It shows that the adaptive subflow allocation algorithm improves the MPTCP throughput by 57% compared to the existing subflow management strategies.
The rest of the paper is organized as follows. Section 2 illustrates observations on MPTCP performance by establishing different number of subflows on different network interfaces, which shows the motivation of adaptive subflow management. Section 3 provides theoretical analysis to the performance of MPTCP in wireless networks with multiple subflows. Section 4 proposes an adaptive subflow management algorithm for MPTCP to optimize its performance gains. Section 5 evaluates the performance of the proposed adaptive subflow management algorithm under a wireless network testbed. Section 6 introduces the related works of MPTCP congestion control and subflow management. Finally the paper is concluded in Section 7.
Section snippets
Observation and motivation
We make several observations to the performance of MPTCP by establishing different number of subflows. Our experimental environment is illustrated in Fig. 1. A laptop is equipped with two WiFi interfaces associated with two APs and a 3G network interface connected to a cellular base station. We test the MPTCP throughput for downloading a large file from the multipath-tcp.org website. We conduct two experiments to observe the MPTCP performance as below.
In the first experiment, we activate only
Analysis of MPTCP performance with uncoupled subflows
In this section, we propose theoretical analysis to the benefit of establishing multiple subflows on a network interface. Given the fact that a large number of congestion control mechanisms [12], [13], [14] for MPTCP have been proposed, we only focus on the “uncoupled” congestion control for MPTCP subflows. In uncoupled mode, the subflows make their own decisions independently to adjust their congestion windows according to the additive increase and multiplicative decrease (AIMD) principle. The
Adaptive subflow allocation mechanism
In this section, we discuss the issues of establishing the proper number of MPTCP subflows on each interface for multi-homed devices to maximize the network performance. According to the previous analysis, establishing more than one subflow on an interface can improve the overall throughput. However, when the number of subflows reaches to some value, the throughput improvement is marginal for larger . Besides, establishing more subflows will increase the consumption of system resources (such
Experiment setup
We setup a wireless network testbed to evaluate the performance of the proposed strategy. Fig. 8 illustrates our testbed. It consists 4 APs (TP-LINK TLWDR4310 wireless routers) deployed in the hallway of the department building of our university. The distance between two APs is about 10 m. Among them , and are connected to the Internet directly, while is configured as a relay for to emulate multi-hop wireless communication. Three laptops (Thinkpad T440, Dell XPS13 L322X, and
Related work
Nowadays most smartphones and mobile devices are multi-homed. To make full use of the capacity of multiple interfaces, efforts have been made to simultaneously transfer data over several network connections [1], [2], [3], [4], [5], which mainly focus on developing multi-path transmission protocols without the change of applications. The Equal-Cost Multipath Routing Protocol (ECMP) [1] is an IP layer protocol that enables routing packets along multiple paths of equal costs to achieve load
Conclusion
In this paper, we illustrated that the current implementation of MPTCP with static subflow management strategies is sub-optimal, and the throughput can be improved greatly by adjusting the number of subflows on each network interface dynamically. We presented theoretical analysis on the performance of MPTCP with subflows, which showed that the overall throughput is sensitive to for lossy communication links. Based on the analytical results, we proposed an adaptive subflow allocation
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was partially supported by the National Key R&D Program of China (Grant No. 2017YFB1001801), the National Natural Science Foundation of China (Grant Nos. 61972196, 61832008, 61832005), the Key R&D Program of Jiangsu Province, China (Grant No. BE2018116), the open Project from the State Key Laboratory of Smart Grid Protection and Operation Control “Research on Smart Integration of Terminal-Edge-Cloud Techniques for Pervasive Internet of Things”, the Collaborative Innovation Center of
Wenzhong Li receives his B.S. and Ph.D. degree from Nanjing University, China, both in computer science. He was an Alexander von Humboldt Scholar Fellow in University of Goettingen, Germany. He is now a full professor in the Department of Computer Science, Nanjing University. Dr. Li’s research interests include distributed computing, edge computing, and pervasive computing. He has published over 100 peer-review papers at international conferences and journals, which include INFOCOM, UBICOMP,
References (39)
- et al.
Papadopoulos packet priority assignment for wireless control systems of multiple physical systems
J. Syst. Archit.
(2020) - et al.
Bounding carry-in interference for synchronous parallel tasks under global fixed-priority scheduling
J. Syst. Archit.
(2018) - et al.
Workload-aware harmonic partitioned scheduling for fixed-priority probabilistic real-time tasks on multiprocessors
J. Syst. Archit.
(2019) - et al.
Multipath issues in unicast and multicast next-hop selection
(2000) Stream control transmission protocol
(2007)- et al.
Concurrent multipath transfer using sctp multihoming over independent end-to-end paths
IEEE/ACM Trans. Netw.
(2006) Multipath Aggregation of Heterogeneous Access Networks
(2011)- et al.
Multipath tcp: From theory to practice
- et al.
How hard can it be? designing and implementing a deployable multipath tcp
- Costin Raiciu, Sebastien Barre, Christopher Pluntke, Adam Greenhalgh, Damon Wischik, Mark Handley, Improving datacenter...
A measurement study on multi-path tcp with multiple cellular carriers on high speed rails
Mptcp meets fec: Supporting latency-sensitive applications over heterogeneous networks
IEEE/ACM Trans. Netw.
Stms: Improving mptcp throughput under heterogeneous networks
Design implementation and evaluation of congestion control for multipath tcp
Mptcp is not pareto-optimal: Performance issues and a possible solution
IEEE/ACM Trans. Netw.
Multipath tcp algorithms: Theory and design
SIGMETRICS Perform. Eval. Rev.
On the benefits of applying experimental design to improve multipath tcp
Experimental evaluation of multipath tcp schedulers
Cited by (5)
Editorial to special issue on resource management for edge intelligence
2021, Journal of Systems ArchitectureA Survey of Multi-Transmission Path Optimization
2023, ACM International Conference Proceeding SeriesCMT-SCTP and MPTCP Multipath Transport Protocols: A Comprehensive Review
2022, Electronics (Switzerland)A Novel Analytical Framework Is Developed for Wireless Heterogeneous Networks for Video Streaming Applications
2022, Journal of MathematicsMathematical model for organizing storage and providing access to files in distributed information systems
2021, Journal of Physics: Conference Series
Wenzhong Li receives his B.S. and Ph.D. degree from Nanjing University, China, both in computer science. He was an Alexander von Humboldt Scholar Fellow in University of Goettingen, Germany. He is now a full professor in the Department of Computer Science, Nanjing University. Dr. Li’s research interests include distributed computing, edge computing, and pervasive computing. He has published over 100 peer-review papers at international conferences and journals, which include INFOCOM, UBICOMP, IJCAI, ACM MM, IEEE JSAC, IEEE/ACM ToN, IEEE TPDS, etc. Prof. Li is a member of IEEE and ACM. He was also the winner of the Best Paper Award of ICC 2009 and APNet 2018.
Lingfan Yu is a Computer Science Ph.D. student at New York University advised by Professor Jinyang Li. He is broadly interested in distributed system research. His primary research work focuses on applying system techniques to improving the performance of machine learning frameworks. Before going to NYU, Lingfan Yu received his B.S. Degree in Computer Science at Nanjing University, China, where he was advised by Dr. Wenzhong Li, and worked on improving performance of wireless networks.
Chaojing Xue received her B.S. degree in Computer Science from Jinling Institute of Technology, China, and received her M.S. degree in Computer Science from Nanjing University, China. She is now working as an engineer in the Huawei Company, Hangzhou, China. Her research interests include multipath TCP, network congestion control, wireless network management, and data mining.
Han Zhang received his B.S. degree in Computer Science from Nanjing University, China. He is now a Master student in the Department of Computer Science, Nanjing University. His research interests include multipath TCP, network congestion control, and reinforcement learning.
Jixiang Lu received his master degree of computer science from Nanjing University, China, 1999. He is now working as a staff in the State Key Laboratory of Smart Grid Protection and Operation Control, NARI Group Corporation. His research interests include artificial intelligence and economic power dispatch.
Rongrong Cao received her master degree of control theory and control engineering from Nanjing University of Science and Technology, China, 2005. She is now working as a staff in the State Key Laboratory of Smart Grid Protection and Operation Control, NARI Group Corporation. Her research interests include smart grid dispatching and control.
Sanglu Lu received her B.S., M.S., and Ph.D. degrees from Nanjing University in 1992, 1995, and 1997, respectively, all in computer science. She is currently a professor in the Department of Computer Science and Technology and the deputy director of State Key Laboratory for Novel Software Technology. Her research interests include distributed computing, pervasive computing, and wireless networks. She has published more than 100 papers in referred journals and conferences in the above areas. She is a member of IEEE and ACM.