TCP-FIT: An improved TCP algorithm for heterogeneous networks☆

https://doi.org/10.1016/j.jnca.2016.03.020Get rights and content

Abstract

Wireless networks and large bandwidth delay product (BDP) networks are two types of challenging environments for TCP congestion control. Many congestion control algorithms have been proposed to improve the performance of TCPs in these two environments. Although these improved algorithms can achieve remarkable performance enhancements for either of the two environments, designing a congestion algorithm that performs well over heterogeneous networks that contain both wireless and large BDP links remains a great challenge. In this study, we propose a novel congestion avoidance algorithm called TCP-FIT, which can perform excellently over both wireless and large BDP links while maintaining good fairness with the standard TCP Reno algorithm. To evaluate the proposed algorithm, theoretical analysis is presented for equilibrium, network utilization, stability, TCP friendliness, RTT fairness, and responsiveness. A series of experimental results obtained using network emulators and over the “live” Internet are also presented to demonstrate the significant performance and fairness improvements of TCP-FIT compared with other state-of-the-art algorithms.

Introduction

The Transmission Control Protocol (TCP) is a reliable transport layer protocol that is widely used on the Internet. Congestion control is an integral module of TCP that directly determines the performance of the protocol. The standard TCP congestion control framework includes four algorithms: “congestion avoidance”, “slow start”, “fast retransmit” and “fast recovery”, among which congestion avoidance plays a crucial role in the performance of long-term TCP flows because it governs the steady-state behavior of the TCP congestion control window. TCP Reno (Jacobson, 1995), the de facto standard TCP algorithm for Internet congestion control, uses the additive increase and multiplicative decrease (AIMD) algorithm in its congestion avoidance, which achieved great success for several decades but has been found to perform poorly over wireless (Tian et al., 2005, Ghaffari, 2015) and large bandwidth delay product (BDP) links (Floyd,, Kushwaha and Gupta, 2014). To improve the performance of TCP Reno over wireless and large BDP links, a substantial number of new TCP variants have been proposed, including TCP Westwood (Mascolo et al., 2001), and TCP Veno (Fu and Liew, 2003) for wireless applications and HighSpeed TCP (Floyd, 2003), Compound TCP (Tan et al., 2006), TCP CUBIC (Ha et al., 2008), FAST TCP (Wei et al., 2006) and HCC (Xu et al., 2011) for large BDP links.

Congestion control algorithms for high-speed large BDP links and high-loss wireless links are commonly considered as two separate research topics requiring different design philosophies and methodologies, leading to various algorithms focusing on only one of the two types of networks but not both. For example, TCP Veno and TCP Westwood can enhance TCP performance over high-loss wireless links but cannot fully adapt to the rapid growth of BDPs in emerging high-speed long-delay networks. HighSpeed TCP, Compound TCP, and TCP CUBIC achieve remarkable throughput improvements in large BDP links but cannot handle random packet losses of wireless links. However, with the deployment of high-speed long-distance networks, such as intercontinental optical networks and advanced high-speed wireless networks such as 4G LTE, the Internet has become increasingly heterogeneous. Therefore, designing a TCP congestion avoidance algorithm that can perform excellently over both large BDP and wireless connections becomes essential for current Internet congestion control.

In this paper, a novel TCP congestion avoidance algorithm, named TCP-FIT, for both large BDP and wireless links is proposed. Theoretical analysis from the equilibrium, network utilization, stability, TCP friendliness, RTT fairness, and responsiveness perspectives and extensive experimental results obtained using both network emulators as well as over the PlanetLab WAN testbeds show that TCP-FIT significantly improved TCP throughput over large BDP, packet loss, and RTT fluctuation environments compared with other state-of-the-art algorithms while maintaining good TCP friendliness, RTT fairness and responsiveness.

The rest of this paper is organized as follows. Section 2 gives an overview of existing TCP variants. Section 3 describes the TCP-FIT algorithm in detail. The throughput model of TCP-FIT is introduced in Section 4, and Section 5 provides theoretical analysis. Performance measured over network emulators and PlanetLab is given in 6 Experimental results, 8 Conclusions. Section 8 concludes the paper.

Section snippets

Related studies

Since the introduction of the first widely used TCP congestion control algorithm in Jacobson (1995), many TCP congestion control algorithms have been proposed. Table 1 contains a list of such algorithms widely found in mainstream operating systems, i.e., Windows and Linux. At a high level, these algorithms can be classified into three categories based on their input signals: namely, loss-based TCP, delay-based TCP and hybrid TCP.

Loss-based TCP includes TCP Reno, TCP Bic (Xu et al., 2004), TCP

Motivation

We define a Packet Loss Event (PLE) as a series of packet losses that starts when the first packet loss is detected by three duplicate ACKs in a window (which triggers the TCP Fast Recovery mechanism), and ends when the last lost packet recovered using Fast Recovery in the same window. The standard AIMD algorithm for adjusting the window w isEachRTT:w←w+1,EachPLE:w←w−w2.Here, we ignore losses that are detected by timeouts because w in timeout is not controlled by Congestion Avoidance. To

Throughput model of TCP-FIT

In this section, we introduce an approximate throughput model for TCP-FIT following the notations of Padhye's TCP Reno throughput model (Padhye et al., 1998). In the model, we represent the AIMD algorithms in a general form, i.e.,EachRTT:w←w+a,EachLoss:w←w−b·w.Obviously a=N, b=23N+1 for TCP-FIT, and a=1, b=1/2 for TCP Reno. Fig. 2 illustrates the evolution of w when w is controlled by the generic AIMD algorithm in (9). The notations in Fig. 2 are summarized in Table 2.

We define the period

Network model

We study the efficiency of TCP-FIT in a widely used network model (Tang et al., 2007). The network model consist of L links, denoted by l∈{1,…,L}, which are shared by S TCP sessions, denoted by s∈{1,…,S}. The routing matrix R of the model is an L×S matrix, where Rl,s={1sessionstraverselinkl0otherwise.The notations used in the network model are summarized in Table 3. Among the notations, Dl and Pl are inherent to the link and are independent of congestion. Oppositely, ql and pl have a closely

Experimental results

In this section, we tested the performance of TCP-FIT over large BDP, random packet loss, random RTT fluctuation, TCP Reno friendliness, RTT fairness and responsiveness scenarios. Because there are so many TCP algorithms and diverse algorithms have different fortes for different scenarios, benchmark selection becomes a very difficult task in our experiments. To find a reasonable benchmark, we tested all the algorithms listed in Table 1 and selected the algorithm with the best performance in the

Experiments over real networks

To investigate the performance of TCP-FIT over wide area networks, we used PlanetLab (Chun et al., 2003) as a testbed to compare TCP-FIT with TCP Reno, CUBIC, and TCP Vegas. In our experiments, 245 PlanetLab nodes located in 192 cities of 43 countries were used. These nodes covered 233 ISPs, representative of the current conditions of the Internet. In our experiments, the PlanetLab nodes simultaneously and repeatedly downloaded seven video clips from an HTTP server located in San Diego,

Conclusions

In this study, we describe a novel TCP congestion control algorithm for application over heterogeneous networks that contain both large BDP and wireless links. Theoretical and emulation results as well as performance measured using live real-world networks show significant performance improvements in the throughput, fairness responsiveness and robustness over state-of-the-art TCP variants.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (Grant nos. 61572059, 61202426), the State Key Program of National Natural Science of China (Grant no. 71531001), the Basic Science Research Program of Shenzhen (Grant no. JCYJ20140828113329138) and the Open Research Fund Program of State Key Laboratory of Software Development Environment (Grant no. SKLSDE-2015ZX-19).

References (47)

  • Chen, M., 2006. A general framework for flow control in wireless networks (Ph.D. thesis). University of...
  • Chen, M., Zakhor, A., 2006. Flow control over wireless network and application layer implementation. In: Proceedings of...
  • B. Chun et al.

    Planetlaban overlay testbed for broad-coverage services

    ACM SIGCOMM Comput. Commun. Rev.

    (2003)
  • J. Crowcroft et al.

    Differentiated end-to-end Internet services using a weighted proportional fair sharing TCP

    ACM SIGCOMM Comput. Commun. Rev.

    (1998)
  • Cui, T., Andrew, L., 2010. FAST TCP for ns-2. URL...
  • T. Flach et al.

    Reducing web latency: the virtue of gentle aggression

    ACM SIGCOMM Comput. Commun. Rev.

    (2013)
  • Floyd, S., 2003. HighSpeed TCP for Large Congestion Windows, RFC...
  • Floyd, S., Allman, M., 2008. Comments on the Usefulness of Simple Best-effort Traffic, RFC...
  • C. Fu et al.

    TCP VenoTCP enhancement for transmission over wireless access networks

    IEEE J. Sel. Areas Commun.

    (2003)
  • Grieco, L., Mascolo, S., 2005. Mathematical analysis of Westwood+ TCP congestion control. In: Proceedings of IEE...
  • S. Ha et al.

    CUBIC: a new TCP-friendly high-speed TCP variant

    ACM SIGOPS Oper. Syst. Rev.

    (2008)
  • V. Jacobson

    Congestion avoidance and control

    ACM SIGCOMM CCR

    (1995)
  • W. Jingyuan et al.

    Improving the incast performance of datacenter TCP by using rate-based congestion control

    IEICE Trans. Fundam. Electron. Commun. Comput. Sci.

    (2014)

    • Cited by (20)

    • A new analytical model of TCP Hybla for satellite IP networks

      2018, Journal of Network and Computer Applications
      Citation Excerpt :

      In this paper we focus particularly on TCP Hybla congestion control. A number of other mechanisms, like INVS (Wang et al., 2016a) and TCP-FIT (Wang et al., 2016b), etc. have been proposed as TCP congestion control for heterogeneous networks. INVS aims at improving network performance over lossy links, whereas TCP-FIT aims at yielding high throughput for high bandwidth-delay product (BDP) networks compared with TCP Reno, Westwood, TCP Vegas etc.

    • Optimized device centric aggregation mechanisms for mobile devices with multiple wireless interfaces

      2017, Computer Networks
      Citation Excerpt :

      Currently, state-of-the-art smartphones do not take advantage of the coexistence of WiFi and cellular interfaces; only one interface can be active at a time, with priority normally given to WiFi irrespective of link qualities. Most proposed techniques in the literature require modifications, or at least some sort of support, at the already-mature layers of the network protocol stack [24,25]. In addition, many require the deployment of a proxy server between mobile devices and destination content servers; in these approaches, the traffic is routed through an intermediate proxy server that implements the proposed bandwidth aggregation support, and that communicates with the destination servers, e.g., see [26].

    • Elephants Sharing the Highway: Studying TCP Fairness in Large Transfers over High Throughput Links

      2023, ACM International Conference Proceeding Series

    • On Minimizing TCP Traffic Congestion in Vehicular Internet of Things (VIoT)

      2023, Wireless Personal Communications

    • A comparative performance analysis of TCP congestion control algorithm: elastic TCP vs. e-Elastic TCP

      2023, International Journal of Internet Protocol Technology

    • TCP-WBQ: a backlog-queue-based congestion control mechanism for heterogeneous wireless networks

      2022, Scientific Reports
    View all citing articles on Scopus
    ☆

    This paper is an extent version of Wang et al. (2011).

    View full text
    © 2016 Elsevier Ltd. All rights reserved.