Elsevier

Computer Networks

Volume 57, Issue 8, 4 June 2013, Pages 1838-1852
Computer Networks

Rethinking the Low Extra Delay Background Transport (LEDBAT) Protocol

https://doi.org/10.1016/j.comnet.2013.02.020Get rights and content

Abstract

BitTorrent has recently introduced LEDBAT, a novel application-layer congestion control protocol for data exchange. The protocol design assumes that network bottlenecks are at the access of the network, and that thus user traffic competes creating self-induced congestion. To relieve this phenomenon, LEDBAT is designed to quickly infer when self-induced congestion is approaching (by detecting relative changes of the one-way delay in the transmission path), and to react promptly by reducing the sending rate prior to the congestion occurrence. Previous work has however shown LEDBAT to be affected by a latecomer advantage, where newly arriving connections can starve already existing flows. In this work, we propose modifications to the congestion window update mechanism of LEDBAT that solve this issue, thus guaranteeing intra-protocol fairness and efficiency. Closed-form expressions for the stationary throughput and queue occupancy are provided via a fluid model, whose accuracy is confirmed by means of ns2 packet level simulations. Our results show that the proposed change can effectively solve the latecomer issue, furthermore without affecting the other original LEDBAT goals.

Introduction

As recently pointed out in [9], “Internet delays now are as common as they are maddening”. The root cause for these delays can be identified with the excess buffering inside a network, which is nicknamed “bufferbloat”. Though this is nothing new [10], the situation got worse in the latest years due to mainly two facts: (i) TCP loss-based design, that forces the bottleneck buffer to fill before the sender reduces his rate and (ii) the fact that low uplink capacities of widely deployed ADSL and Cable modems can translate into significant queuing delay (up to few seconds [24]).

Evidently, BitTorrent engineers were well aware of this fact. Indeed, the popular peer-to-peer file sharing system with hundreds of millions of daily users worldwide, has recently developed a novel application-layer congestion control protocol for data exchange. The novel insight in the widely explored congestion control landscape is in this case the reasonable assumption that the bottleneck is most likely at the access of the network (e.g., at the ADSL modem line), which means that congestion is therefore typically self-induced by concurrent traffic sessions generated by the same user (e.g., BitTorrent transfers in parallel with Skype call and Web browsing). The new protocol, named LEDBAT after Low Extra Delay Background Transport, is designed to solve this issue and targets (i) efficient but (ii) low priority transfers. When LEDBAT flows have the exclusive use of the bottleneck resources, they fully exploit the available capacity. When instead other transfers – such as VoIP, gaming, Web or other TCP flows – are ongoing, LEDBAT flows back off to avoid harming the performance of interactive traffic. To attain the efficiency aim, LEDBAT flows need to create queuing, as otherwise the capacity would not be fully utilized. At the same time, due to the low-priority aim, the amount of extra queuing delay induced by LEDBAT flows should be small enough to avoid hurting the interactive traffic – hence the protocol name.

LEDBAT2 has been defined as an IETF draft [37] (which focuses more on the algorithmic aspects) and as a BitTorrent Enhancement Proposal BEP-29 [32] (that instead focuses more on the UDP framing). and has recently become the BitTorrent default congestion control protocol, replacing thus TCP for the data transfer.

While previous research [33], [34], [8], [7], [4], [13], [36], [39], [40] on LEDBAT has shown its potential for P2P transfers it has also highlighted some serious fairness deficiencies. In more details, LEDBAT has a good interplay with BitTorrent-like transmissions of short-lived chunk-long flows [39], [40]: hence, the definition of this protocol under a BEP document makes perfectly sense, as the performance improvement for BitTorrent is in this case coupled to a reduction of the self-induced bufferbloat. At the same time, LEDBAT is affected by a latecomer unfairness issue [34], [8] that arises in (the not so uncommon) case of backlogged flows: under this conditions, latecomer flow takes over the bottleneck resource, starving the first-comers. As such, the normalization of an ill-defined protocol can have potentially dramatic effects: while the phenomenon hardly ever happens in P2P swarm like content delivery, it can severely impact the performance of backups, photo uploads and, more generally, long lived uploads of home users towards their virtual storage (e.g. DropBox, Google, Windows Live, Apple iCloud, etc.).

With respect to the normalization of an IETF congestion control algorithm, whose scope goes beyond a specific application, though popular it may be, fairness is a significant property that should always be taken into account for the design of any experimental or deploy-able protocol. The lack of fairness is an indication of poor convergence properties, e.g. the algorithm is unstable because of an unwise choice of the parameters or, more tricky, the algorithm cannot be stable for any choice of parameter. We show in this paper that LEDBAT falls in this second category.

The main contribution of this work is to propose a modification to the LEDBAT congestion control that, leaving untouched the design goals, solves the fairness issue – therefore avoiding unwanted effects at the application-layer altogether. Throughout this paper, we make use of several complementary techniques to study our proposal. First, we use passive measurements to gauge the popularity of LEDBAT transfers in the real Internet, and exploit an active testbed methodology to show the fairness issue in current BitTorrent. We propose a modification to the original LEDBAT protocol, to jointly achieve fairness and efficiency, and develop a fluid model to describe the system dynamics. An analytical solution of the model proves the soundness of our design, while numerical solutions allow us to grasp the transient phase as well. Finally, we use extensive ns2 packet-level simulations to evaluate the LEDBAT performance under several scenarios. Since we need to ensure that our proposal works as expected under any circumstances, we include the general case of backlogged transfers (e.g., two concurrent low priority backups). At the same time, we need to ensure that LEDBAT does not harm the experience of BitTorrent users, hence we include P2P-like scenarios involving multiple-flows and a heterogeneous network environment as well.

The remainder of this paper is organized as follows. Related work and motivations are covered in Sections 2 Related work, 3 Motivation respectively. The unfairness issue is introduced in Section 4, followed by our proposed modification in Section 5 along with the fluid model and its analytical solution. More complex network scenarios are tackled by means of packet-level simulations starting from Section 6, where we also compare the numerical solution of the fluid model with simulation results. We then proceed by studying the impact of the traffic model (e.g., backlogged vs. chunk-based transfers) in Section 7, the sensitivity of the protocol to parameter changes in Section 8. Lastly we evaluate the performance of the protocol in P2P-like settings from a single-peer as well as a whole swarm perspective in Section 9, before Section 10 concludes the paper.

Section snippets

Related work

Congestion control studies on the Internet date back to [21] and it is out-of-scope to provide a full review of the existing literature here. Still, a couple of references are worth citing as they share LEDBAT low-priority spirit [41], [25], [28], [23]. For instance, TCP-LP [25] and NICE [41] share the same goal as LEDBAT aiming at implementing a Lower-than Best Effort (LBE) service for background transfers. In more detail, NICE extends the delay-based behavior, typical of TCP Vegas, with a

Motivation

Despite recent research showing an increasing importance of video over the share of Internet traffic [3], BitTorrent still represents a significant portion of user generated data – and due to the recent shutdown of popular file-hosting services such as Megaupload/Megavideo [15], we can expect the Internet ratio of file-sharing to increase again. In Fig. 1a, we depict, over the last few years, the BitTorrent traffic share (UDP and TCP traffic, overall traffic) averaged at the PoP of 5 European

Current LEDBAT fairness issues

According to the original draft proposal [37], LEDBAT maintains a minimum One-Way Delay (OWD) estimation Dmin, which is used as base delay to infer the amount of delay due to queuing. LEDBAT flows have a target queuing delay τ, i.e., they aim at introducing a small, fixed, amount of delay in the queue of the bottleneck buffer. Flows monitor the variations of the queuing delay q(t)  Dmin to evaluate the distance Δ(t) from the target:Δ(t)=(q(t)-Dmin)-τ,where q(t) is the queueing delay measured at

Proposed LEDBAT modification

To address the latecomer issue, we propose to modify the delay-based decrease term and to introduce a multiplicative decrease continuously driven by the estimated distance from the target, Δ(t). Intuitively, the multiplicative window reduction response to congestion allows the source sending rate to slow down enough to make a stable (and fair) point always reachable. Clearly, to guarantee at the same time fairness and protocol efficiency, a proper choice of the decrease factor has to be made,

Simulation overview

So far we have developed a mathematical model of our new proposed protocol in order to formally prove its properties. However, the model is based on a number of simplifying assumptions and it neglects some aspects due to packet-level quantization (i.e., queue length and congestion window in multiple of fixed-size packets as opposed to continuous rate in the fluid model). Hence, in the remainder of this work we carry out a thorough packet-level ns2 [2] simulation campaign, to cope with scenarios

Impact of traffic model

In this section we assess how the fLEDBAT protocol deals with different kinds of traffic. Besides the classical backlogged transfer, we simulate a chunk-based transfer, which mimics the behavior of a BitTorrent data exchange between two peers.

Sensitivity analysis

In this section we carry out further simulations to assess the impact of the choice of the parameter ζ on the protocol performance. In order to gather a complete sensitivity analysis of fLEDBAT parameters, we consider several scenarios: (i) a TCP NewReno flow competing with a fLEDBAT flow, (ii) a LEDBAT flow competing with a fLEDBAT flow, (iii) two or more fLEDBAT flows competing at the same bottleneck. All flows operate in chunk-by-chunk transmission mode.

As performance metrics, we consider

P2P scenarios

In order to compare fLEDBAT vs. LEDBAT performance under other conditions, closer to the BitTorrent use-case, we finally consider a chunk-based scenario that (i) loosely mimics the behavior of BitTorrent peers, (ii) employs Internet-like heterogeneous delays and access rates, (iii) considers background traffic and coupled queues, (iv) addresses the impact of chunk sizes.

We consider two different scenarios: first a single BitTorrent peer and a single queue in isolation, using a more

Conclusion

In this work, we proposed modifications to the LEDBAT congestion control algorithm that not only are able to achieve low-priority inter-protocol (i.e., against TCP) and efficiency intra-protocol (e.g., with other fLEDBAT flows), but also reintroduce intra-protocol fairness, solving thus the late-comer issues of the original LEDBAT proposal.

To model the protocols dynamic we used a fluid-model approach which allows, on the one hand, to detect the main issue at the base of LEDBAT unfairness (i.e.,

Giovanna Carofiglio received the M.Sc. degree from Politecnico di Turing, Italy, in 2004, and the P.D. from Politecnico di Torino and Telecom ParisTech, Paris, France, in 2008. Her graduate research focused on stochastic analysis of wired and wireless networks and has been performed at Politecnico di Torino and at Ecole Normale Superieure (ENS Ulm) in the INRIA-TREC group. Since July 2008, she is with Alcatel-Lucent Bell Labs, Paris, working on design and performance evaluation of communication

References (44)

  • D.M. Chiu et al.

    Analysis of the increase and decrease algorithms for congestion avoidance in computer networks

    Comput. Netw. ISDN Syst.

    (1989)
  • S. Liu et al.

    TCP-Illinois: a loss-and delay-based congestion control algorithm for high-speed networks

    ACM Perform. Eval.

    (2008)
  • LEDBAT ns2 code....
  • The Network Simulator ns2....
  • Cisco visual networking index: forecast and methodology, 2008–2013, June 2009....
  • A.J. Abu, S. Gordon, A dynamic algorithm for stabilising LEDBAT congestion window, in: 2nd IEEE International...
  • Catherine Boutremans, Le Boudec, Jean-Yves, A note on the fairness of TCP vegas, in: International Zurich Seminar on...
  • L.S. Brakmo et al.

    TCP vegas: new techniques for congestion detection and avoidance

    ACM SIGCOMM Comput. Commun. Rev.

    (1994)
  • G. Carofiglio, L. Muscariello, D. Rossi, C. Testa, A hands-on Assessment of Transport Protocols with Lower than Best...
  • G. Carofiglio, L. Muscariello, D. Rossi, S. Valenti, The quest for LEDBAT fairness, in: IEEE Global Communication...
  • Vint Cerf et al.

    Bufferbloat: what’s wrong with the internet?

    Commun. ACM

    (2012)
  • Stuart Cheshire, It’s the latency, stupid!, May 1996....
  • B. Cohen, Incentives build robustness in BitTorrent, in: Workshop on Economics of Peer-to-Peer systems, Cambridge, MA,...
  • B. Cohen, A. Norberg, Correcting for clock drift in uTP and LEDBAT, in: Invited talk at 9th USENIX International...
  • A. Finamore et al.

    Experiences of internet traffic monitoring with tstat

    IEEE Network Magazine, Special Issue on Network Traffic Monitoring and Analysis

    (2011)
  • Franich, D. Megaupload (and megavideo) shut down by the feds, January 2012....
  • Qing Gao et al.

    Adaptive vegas: a solution of unfairness problem for tcp vegas

  • J. Garcia-Dorado, A. Finamore, M. Mellia, M. Meo, M. Munafó, Characterization of ISP traffic: trends, user habits, and...
  • G. Hasegawa et al.

    Fairness and stability of congestion control mechanisms of tcp

    Telecommun. Syst.

    (2000)
  • G. Hazel, uTorrent Transport Protocol library, May 2010....
  • S. Hemminger et al., Network emulation with NetEm, in: Linux Conf Australia (LCA 2005), Canberra, Australia, April...
  • V. Jacobson, Congestion avoidance and control, in: ACM SIGCOMM Comp. Comm. Rev., Stanford, CA, August...
  • Cited by (0)

    Giovanna Carofiglio received the M.Sc. degree from Politecnico di Turing, Italy, in 2004, and the P.D. from Politecnico di Torino and Telecom ParisTech, Paris, France, in 2008. Her graduate research focused on stochastic analysis of wired and wireless networks and has been performed at Politecnico di Torino and at Ecole Normale Superieure (ENS Ulm) in the INRIA-TREC group. Since July 2008, she is with Alcatel-Lucent Bell Labs, Paris, working on design and performance evaluation of communication networks. She served as TPC member of several conferences including IEEE INFOCOM.

    Luca Muscariello received the M.Sc. and Ph.D. degrees from Politecnico di Torino, Italy, in 2002 and 2006 respectively. His graduate research was performed in the area of Internet traffic measurement, characterization and modeling at Politecnico di Torino, France Telecom R&D in Paris and at VTT in Helsinki. Since 2006, he is R&D engineer at Orange Labs, France Telecom R&D, Paris, France. His current research interests include simulation and analytical modeling of wired and wireless networks. He served as TPC member of several conferences including IEEE INFOCOM.

    Dario Rossi received the M.Sc. and Ph.D. degrees from Politecnico di Torino, Italy in 2001 and 2005 respectively. During 2003/04 he was with the CS Department at University of California, Berkeley. Since October 2006, he is an Associate Professor at Telecom ParisTech, Paris, France. He has coauthored over 70 papers in leading conferences and journals, holds 4 patents and he participated in the program committees of over 30 conferences including IEEE INFOCOM, ICC, IPCCC and GLOBECOM. His research interests include green networking, peer-2-peer networks, Internet traffic measurement and traffic engineering.

    Claudio Testa received the M.Sc. degree in computer and communication networks engineering from the Politecnico di Torino, Italy, discussing his Master Thesis in the network traffic analysis field. Since April, he is a Ph.D. student at the Computer Science and Networking department of Telecom ParisTech, Paris, France. His research interest are in the field of congestion control of peer-to-peer services, network measurement and network traffic characterization.

    Silvio Valenti (S’08) received the M.Sc. degree in computer science engineering from Politecnico di Torino, Italy, in February 2008, on the development and optimization of a cross-compiler for network processor architectures. Since August 2008, he is a Ph.D. student at the Computer Science and Networking department at Telecom ParisTech, in Paris, France. His research interests consists in peer-2-peer networking, Internet traffic classification and high-speed packet processing.

    1

    Current affiliation: Google Inc.

    View full text