Adaptive subflow allocation for multipath data transmission in mobile edge networks

Abstract

Multipath data transmission techniques have been proposed to aggregate the resource of multiple heterogeneous access networks to enhance the robustness and throughput of multi-homed mobile devices. The Multipath TCP (MPTCP) protocol has been standardized by the IETF as an extension of conventional TCP, which enables establishing several subflows over multiple interfaces for simultaneous data transmission. However, current implementation of MPTCP only supports static subflow management strategies, which cannot adapt to the changing to network environment. In this paper, we focus on the issues of dynamic subflow management for MPTCP. We present theoretical analysis to the overall throughput of establishing multiple subflows with MPTCP, based on which we propose an adaptive subflow allocation algorithm to determine the proper number of subflows for each network interface to optimize the performance gain. We implement the proposed algorithm, integrate it into the MPTCP Linux Kernel implementation, and evaluate its performance on a wireless network testbed. Extensive experiments show that the proposed allocation algorithm significantly outperforms the existing subflow management strategies with 57% improvement.

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 $n$ 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.