All Internet routers contain buffers to hold packets during times of congestion. Today, the size of the buffers is determined by the dynamics of TCP's congestion control algorithm. In particular, the goal is to make sure that when a link is congested, it is busy 100% of the time; which is equivalent to making sure its buffer never goes empty. A widely used rule-of-thumb states that each link needs a buffer of size B = RTT C, where RTT is the average round-trip time of a flow passing across the link, and C is the data rate of the link. For example, a 10Gb/s router linecard needs approximately 250ms 10Gb/s = 2.5Gbits of buffers; and the amount of buffering grows linearly with the line-rate. Such large buffers are challenging for router manufacturers, who must use large, slow, off-chip DRAMs. And queueing delays can be long, have high variance, and may destabilize the congestion control algorithms. In this paper we argue that the rule-of-thumb (B = RTT is now outdated and incorrect for backbone routers. This is because of the large number of flows (TCP connections) multiplexed together on a single backbone link. Using theory, simulation and experiments on a network of real routers, we show that a link with n flows requires no more than B = (RTT # n, for long-lived or short-lived TCP flows. The consequences on router design are enormous: A 2.5Gb/s link carrying 10,000 flows could reduce its buffers by 99% with negligible difference in throughput; and a 10Gb/s link carrying 50,000 flows requires only 10Mbits of buffering, which can easily be implemented using fast, on-chip SRAM.

In networks that support Quality of Service (QoS), an admission control algorithm determines whether or not a new traffic flow can be admitted to the network such that all users will receive their required performance. Such an algorithm is a key component of future multi-service networks as it determines the extent to which network resources are utilized and whether the promised QoS parameters are actually delivered. Our goals in this paper are threefold. First, we describe and classify a broad set of proposed admission control algorithms. Second, we evaluate the accuracy of these algorithms via experiments using both on-off sources and long traces of compressed video; we compare the admissible regions and QoS parameters predicted by our implementations of the algorithms with those obtained from trace-driven simulations. Finally, we identify the key aspects of an admission control algorithm necessary for achieving a high degree of accuracy and hence a high statistical multiplexing gain...

... In this paper we continue the development of a modelling approach which attempts to integrate these several time-scales, and illustrate its application to the analysis of a family of simple and robust measurement-based admission controls. A subsidiary aim of the paper is to shed light on the relationship between the admission control proposed for ATM networks by Gibbens et al [9] and that proposed for controlled-load Internet services by Floyd [7]. We shall see that their common origin in Chernoff bounds allows the definition of a simple and general family of admission controls, capable of tailoring for several implementation scenarios.

Statistical sharing over several time-scales is a key feature of the Internet, and is likely to be an essential aspect of future ATM networks.

The Differentiated Services framework (DiffServ) has been proposed to provide multiple Quality of Service (QoS) classes over IP networks. A network supporting multiple classes of service also requires a differentiated pricing structure. We propose a pricing scheme in a DiffServ environment based on the cost of providing different levels of quality of service to different classes, and on long-term demand. Pricing of network services dynamically based on the level of service, usage, and congestion allows a more competitive price to be offered, allows the network to be used more efficiently, and provides a natural and equitable incentive for applications to adapt their service contract according to network conditions. We develop a DiffServ simulation framework to compare the performance of a network supporting congestion-sensitive pricing and adaptive service negotiation to that of a network with a static pricing policy. Adaptive users adapt to price changes by adjusting their sending rate or selecting a different service class. We also develop the demand behavior of adaptive users based on a perceptually reasonable user utility function. Simulation results show that a congestion-sensitive pricing policy coupled with user rate adaptation is able to control congestion and allow a service class to meet its performance assurances under large or bursty offered loads, even without explicit admission control. Users are able to maintain a stable expenditure. Allowing users to migrate between service classes in response to price increases further stabilizes the individual service prices. When admission control is enforced, congestion-sensitive pricing still provides an advantage in terms of a much lower connection blocking rate at high loads. I.

Abstract. In this paper, we undertake the first study of statistical multiplexing from the perspective of approximation algorithms. The basic issue underlying statistical multiplexing is the following: in high-speed networks, individual connections (i.e., communication sessions) are very bursty, with transmission rates that vary greatly over time. As such, the problem of packing multiple connections together on a link becomes more subtle than in the case when each connection is assumed to have a fixed demand. We consider one of the most commonly studied models in this domain: that of two communicating nodes connected by a set of parallel edges, where the rate of each connection between them is a random variable. We consider three related problems: (1) stochastic load balancing, (2) stochastic bin-packing, and (3) stochastic knapsack. In the first problem the number of links is given and we want to minimize the expected value of the maximum load. In the other two problems the link capacity and an allowed overflow probability p are given, and the objective is to assign connections to links, so that the probability that the load of a link exceeds the link capacity is at most p. In binpacking we need to assign each connection to a link using as few links as possible. In the knapsack problem each connection has a value, and we have only one link. The problem is to accept as many

Simple and robust engineering rules for dimensioning bandwidth for elastic data traffic are derived for a single bottleneck link via normal approximations for a closed-queueing-network (CQN) model in heavy traffic. Elastic data applications adapt to available bandwidth via a feedback control such as the Transmission Control Protocol (TCP) or the Available Bit Rate transfer capability in ATM. The dimensioning rules satisfy a performance objective based on the mean or tail-probability of the per-flow bandwidth. For the mean objective we obtain a simple expression for the effective bandwidth of an elastic source. We provide a new derivation of the normal approximation in CQNs using more accurate asymptotic expansions and give an explicit estimate of the error in the normal approximation. A CQN model was chosen to obtain the desirable property that the results depend on the distribution of the file sizes only via the mean, and not the heavy-tail characteristics. We view the exogenous "load...

This paper describes methods of computing usage charges from simple measurements and relating these to bounds on the effective bandwidth. Thus we show that charging for usage on the basis of effective bandwidths can be well-approximated by charges based on simple measurements. Charging and pricing are essential requirements in the operation of a communication network. They are needed not only to recover costs and make a profit. Even if a generous operator is willing to offer a network for free, there are still compelling reasons to charges for services in order to exercise control. The congestion that has plagued the Internet because it lacks any mechanism for charging and pricing highlights the fact that without charges it is difficult to control congestion or divide network resources amongst users in a workable and stable way. Subject classifications: Communications: measurement-based charging. Of course there are many considerations that influence the prices at which an operator will choose to sell network services. Marketing and regulation are certainly important, but these considerations are not unique to the operation of a communications network. Special considerations do, however, arise from the fact that a broadband communications network is intended simultaneously to carry a wide variety of traffic types. Our conception of a broadband network is that of a collection of resources (links, buffers, switches, etc.) which can be used to provide a wide variety of communications services. These services are distinguished by traffic contracts, which specify parameters to which the traffic must adhere (a maximum peak rate, for example), and the quality of service which the network undertakes to guarantee (typically, cell loss or delay). These concepts are accepted as ...

: We present an approach to manage and price service level agreements (SLAs) for differentiated services that uses a simple upper bound for the effective bandwidth of the conforming traffic as a proxy for resource usage. The bound depends on the user's traffic profile (peak rate and token bucket descriptor). Usage charges for a specific time period are proportional to this proxy, and their calculation requires only measurements of volume. We discuss and present experimental results regarding the incentives and fairness of the proxy, which is required in order to achieve economic efficiency. An important feature of our approach is the simplicity of the user's procedure for selecting optimal token bucket parameters. Our approach is quite generic and can be applied to scheduling disciplines that enable the provision of multiple service classes with different levels of performance. Finally, we present a case study for two service classes, real-time and non-real-time, with actual Internet t...

Abstract not found