Trade & Cap: A customer-managed, market-based system for trading bandwidth allowances at a shared link
Introduction
Motivation: The ever increasing appetite for Peer-to-Peer (P2P), media streaming, and Video on Demand (VoD) content is forcing service providers to constantly upgrade their infrastructures to keep-up with customers bandwidth demands. This state-of-affairs is significantly exacerbated by the prevalence of flat-pricing schemes and hence the lack of an incentive for users to moderate their hunger for network bandwidth, especially around periods of peak network utilization, which are the primary determinants of an Internet Service Provider (ISP) costs (both in terms of infrastructure upgrade cycle and inter-AS traffic volume costs du e to the 95/5 rule). Attempts by ISPs to deviate from flat pricing (including field-tested per-byte pricing [1]) have been widely rejected by customers [2]. This is also reinforced by the prevalence of flat pricing in the telephony market [3].
In addition to the significant capital investments that ISPs must shoulder to ensure that their networks are well provisioned during the few hours of peak demand, the new “Internet world order” of seemingly unbounded hunger for bandwidth further complicates fundamental issues that have confounded the networking community for decades, including the adoption of an acceptable notion of fairness as it relates to congestion management. Congestion increases delay and losses, reducing the perceived Quality of Service (QoS) of interactive applications such as web browsing, VoIP, and video streaming. Dealing with congestion requires that users (flows) “pay” for their share of the congestion they cause [4], resulting in a degradation in QoS (the congestion price). But, when interactive applications are forced to compete with non-interactive applications – such as P2P filesharing, background backup services, or VoD downloads – the degradation in QoS becomes unacceptable.
Under flat pricing, during periods of peak demand, current congestion control practices could be seen as particularly “unfair” to users of low-volume, mostly-interactive applications who would be effectively subsidizing “bandwidth hogs.” This has prompted some ISPs to act as arbiters, proactively shaping user traffic by setting quotas,3 or by preferentially treating different traffic payloads (e.g., web browsing vs. bittorrent downloads) during periods of peak demand.4 These efforts have backfired, eliciting a public relations quagmire regarding violation of “Net Neutrality,” [17], [18] which is perceived as the prime reason for the Internet being the cradle of innovation it is [19]. Proactive ISP intervention based on traffic payload also raises concerns regarding monopolistic practices, e.g., blocking or taxing Video/VoIP services not provided by the same ISP [19].
Although QoS mechanisms have been around for a while, their deployment on the Internet is minimal. The wide-spread adoption of flat-rate pricing models prevents the utilization of currency-based mechanisms for assigning priority classes [20], [21], [22]. Other mechanisms that do not rely on currency, such as Re-feedback [23] handle congestion by policing flows on the network, but do not give the end-user the means to expresses preferences between different classes of traffic, and thus make it impossible to guarantee the QoS level for certain applications. In contrast, the T&C mechanism provides the means to trade bandwidth allocations, so that the minimum expected demands are ensured through the provider’s network (or the last mile), where bottlenecks typically materialize. Giving the end-users a guaranteed minimum bandwidth allocation makes it possible to prioritize the traffic right at the edge, where policies are under the control of the end-user and existing mechanisms can be used to provide guarantees. While the general strategy used in T&C could be adopted for handling bandwidth allocations at the core of the network, in this paper we restrict our attention to its use at the edge.
We note that the term “Cap and Trade” is widely known in (and typically associated with) the market mechanism used to control the emission of pollutants into the environment [24]. The mechanism we propose in this paper also uses marketplace supply (or allowances) and demand as a mechanism to control congestion in the network. As our mechanism operates first by trading and then by capping, we call it “Trade & Cap”.
Scope and contributions: Rather than having ISPs act as arbiters who set the rules regarding what constitutes fair usage of a shared resource, in this paper, we propose a market-based T&C system in which user software agents converge on what they perceive as an equitable allocation of resources, irrespective of what these resources are used to support (HTTP vs P2P traffic) and irrespective of the absolute resource allocation (traffic volume) per user.5 In our setting, the role of the ISP is that of providing a mechanism that supports any privately-defined user policy [26].
Effectively, our proposed T&C mechanism sets up a marketplace. Given the fixed (flat-rate) payment to the provider, customers enter this marketplace with equal buying power, but their use of this fairly-allocated buying power depends on their flexibility. This allows customers to trade “volume” during low-utilization periods for “quality” during peak-utilization periods (or vice versa). The direction of the trade (not to mention the user’s willingness to even engage in trading) depends entirely on user preferences and flexibility (e.g., tolerance for delaying a scheduled network backup job).6 In addition to empowering customers to trade bandwidth allocations, T&C has the desirable side effect of smoothing traffic utilization over time, thus reducing the ISP’s cost which is determined primarily by the peak rate.
Outline and summary of results: We start this paper in Section 2 by overviewing the T&C mechanism as it applies to a Digital Subscriber Line Access Multiplexer (DSLAM) setting, and in Sections 3 The bandwidth trading phase, 4 The bandwidth capping phase by presenting analytical results pertaining to convergence and efficiency of the marketplace underlying T&C. Formulating the problem as a game is not only useful for purposes of modeling and understanding the marketplace dynamics, but also it serves as the basis of a real mechanism that can be implemented and applied in practice. Thus, in Section 6 we discuss the salient features of an implementation architecture for T&C in a DSLAM setting. Our implementation allows the marketplace interactions to be carried out by software agents that run on behalf of the users and the ISP, and thus (with the exception of minimal configuration and parametrization) is quite transparent to the end user. Next, in Section 7, we demonstrate the significant advantages of T&C by presenting results from extensive trace-driven simulations. For instance, we show that introducing a relatively small level of flexibility in the scheduling of user activities results in significant gains for both the users and the ISP. For example, allowing user agents to reposition bandwidth allocations within relatively small windows of time enables them to increase their share of fluid bandwidth (supporting non-interactive applications) by 20–40% depending on their flexibility. This benefits the ISP as well, resulting in as much as 16–31% reduction in the 95th percentile of the ISP’s 5-min traffic volume, and (even more impressively) resulting in smoothing traffic volume, reducing the 95th-percentile/50th-percentile ratio from 1.58 to an almost perfect ratio of 1.004. We conclude the paper in Section 8 with a review of the related literature.
Section snippets
T&C in a DSLAM setting
While our T&C mechanism is applicable to any setting in which it is desirable to coordinate the fractional acquisition by a set of rational parties of the shared capacity of a single resource, in this paper, and without loss of generality, we restrict ourselves to a specific setting – that of coordinating the utilization of a shared DSLAM link.
Fig. 1 illustrates the basic architecture of Digital Subscriber Line (DSL) access technology. In this setting, DSL modems on the customer side connect
The bandwidth trading phase
Each agent i represents its RT demand as a vector of requested bandwidth allocations: . An assignment of an agent’s demand is a mapping that pins each one of the components of the vector to a different time slot. A set of such assignments (one per agent) comprises a potential configuration, or schedule of RT utilization at the DSLAM.
Let k = mi(j) be the time slot assigned to the jth component of agent i’s request vector. We denote by xik the actual allocation for agent i in
The bandwidth capping phase
The Capping Phase computes a market-clearing solution that allocates the left-over budget of the agents in such a way that maximizes the aggregate FT allocation for each user. Let Wi = (wi1, … , wiT) be the vector of FT allocations, where is the allocation of FT for agent i in time-slot p. We adjust the definition of the cost function to take into account the allocation of FT as follows: Definition 2 The cost to agent i for the combined allocation of RT (xip) and FT (wip) iswhere
Other application scenarios
Load balancing problems arise in a multitude of situations, of which the DSLAM scenario we have considered so far is but one example. The model we have presented is general and can be applied in other scenarios where the customer tasks can be modeled as a combination of atomic and fluid processes and all the customers compete to complete their tasks with the lowest cost.
An example setting in which T&C is applicable is given by Greenberg et al. [31] – namely provisioning datacenter resources. In
Implementation of a T&C DSLAM marketplace
Architecture: We describe a distributed implementation of the T&C marketplace, where there is one provider agent (running at the DSLAM for example), and a client-side agent running on the customer’s local router. The general architecture of the system is illustrated in Fig. 5. In this architecture, the client-side agent is responsible for: (1) profiling the customer’s RT demand, (2) bidding for allocations during the bandwidth trading phase, and (3) shaping applications’ traffic according to
Experimental evaluation
In this section we use trace-driven simulations to (1) highlight the benefits that a user in our system begets by exhibiting some flexibility in scheduling its RT sessions under T&C, (2) demonstrate the gains that an ISP stands to realize as a result of the overall smoother traffic profile of T&C, and (3) illustrate how various parameters affect the performance of T&C.
Traces and Trace Pre-Processing: As an alternative to direct DSLAM traces (which unfortunately are not available), we used
Related work
While the application of game-theoretic and micro-economic approaches to networking problems is not novel [36], [27], [4], [37], [21], our approach of strategically trading-off allocation slots based on desirable properties for different traffic classes is new and quite promising.
Laoutaris and Rodriguez [5] recognized that the problems associated with rampant FT traffic are due to the lack of incentives for end-users to properly schedule their FT traffic and the lack of network mechanisms to
Conclusion
Trade & Cap is an effective bandwidth management mechanism that enables self-interested user agents to collectively converge on what they perceive to be an equitable allocation, based on their individual, private valuation of network utility (e.g., raw volume vs. QoS over time). T&C not only benefits users by allowing them to extract better utility from the network, but also benefits the ISP by yielding smoother aggregate traffic volumes, which lowers traffic transit costs and reduces the
Jorge Londoño recently obtained his Ph.D. degree from the Computer Science Department at Boston University, Boston, MA. He obtained a MA from Boston University in 1999 and a BS at the Universidad Pontificia Bolivariana in 1992. His research interests include distributed systems, and applications of game-theory and micro-economics to resource management problems in these systems.
References (43)
Internet pricing and the history of communications
Comput. Networks
(2001)- M.H. Bosworth, Time warner: metered broadband will prevent “internet brownouts”,...
- W. Gruener, Time warner shelves metered internet plans – for now,...
Charging and rate control for elastic traffic
Eur. Trans. Telecommun.
(1997)- N. Laoutaris, P. Rodriguez, Good things come to those who (can) wait or how to handle delay tolerant traffic and make...
- et al.
Early application identification
- et al.
BLINC: multilevel traffic classification in the dark
- et al.
Transport layer identification of P2P traffic
- et al.
Internet traffic classification using bayesian analysis techniques
- A.W. Moore, K. Papagiannaki, Toward the accurate identification of network applications, in: Passive and Active Network...
Accurate, scalable in-network identification of P2P traffic using application signatures
A radical new router
IEEE Spectr.
Net neutrality: the technical side of the debate: a white paper
SIGCOMM Comput. Commun. Rev.
Pricing congestible network resources
IEEE J. Sel. Areas Commun.
Analysis of a static pricing scheme for priority services
IEEE/ACM Trans. Networks
Cited by (1)
Tragedy of the commons: Self-interest, at the expense of the system
2012, Annual International Conference of the American Society for Engineering Management 2012, ASEM 2012 - Agile Management: Embracing Change and Uncertainty in Engineering Management
Jorge Londoño recently obtained his Ph.D. degree from the Computer Science Department at Boston University, Boston, MA. He obtained a MA from Boston University in 1999 and a BS at the Universidad Pontificia Bolivariana in 1992. His research interests include distributed systems, and applications of game-theory and micro-economics to resource management problems in these systems.
Azer Bestavros is professor and former chairman of Computer Science at Boston University, which he joined in 1991 after completing his Ph.D. at Harvard University. He is the Chair of the IEEE Computer Society Technical Committee on the Internet and a distinguished speaker of the IEEE. He has received distinguished service awards from both the ACM and the IEEE, and was the recipient of the United Methodist Scholar/Teacher Award in 2010. Funded by more than $15 M of government and industry grants, his research is mostly applied to networking and real-time systems, culminating so far in 14 Ph.D. theses, two startup companies, and over 3,000 citations. Some of his early groundbreaking contributions include stochastic extensions of classical rate-monotonic analysis for real-time systems in the early 1990s, pioneering the CDN push content distribution model in the mid 1990s, and seminal Internet traffic characterization and reference locality modeling in the late 1990s. His more recent research has focused on network transport, caching, and streaming media delivery, adversarial exploits of system dynamics, economics-inspired approaches to resource management in overlay, P2P, and cloud settings, and formal approaches to the design and implementation of safety–critical cyber-physical systems.
Nikolaos Laoutaris is a researcher at the Internet research group of Telefonica Research in Barcelona. Prior to joining the Barcelona lab he was a postdoc fellow at Harvard University and a Marie Curie postdoc fellow at Boston University. He got his Ph.D. in computer science from the University of Athens in 2004.
- 1
Supported in part by the Universidad Pontificia Bolivariana and COLCIENCIAS–Instituto Colombiano para el Desarrollo de la Ciencia y la Tecnología “Francisco José de Caldas”.