While today’s computer networks support only best-effort service, future packet-switching integrated-services networks will have to support real-time communication services that allow clients to transport information with performance guarantees expressed in terms of delay, delay jitter, throughput, and loss rate. An important issue in providing guaranteed performance service is the choice of the packet service discipline at switching nodes. In this paper, we survey several service disciplines that are proposed in the literature to provide per-connection end-to-end peqormance guarantees in packet-switching networks. We describe their mechanisms, their similarities and differences, and the performance guarantees they can provide. Various issues and tradeoffs in designing service disciplines for guaranteed performance service are discussed, and a general framework for studying and comparing these disciplines are presented. I.

Suppose that n balls are placed into n bins, each ball being placed into a bin chosen independently and uniformly at random. Then, with high probability, the maximum load in any bin is approximately log n log log n . Suppose instead that each ball is placed sequentially into the least full of d bins chosen independently and uniformly at random. It has recently been shown that the maximum load is then only log log n log d +O(1) with high probability. Thus giving each ball two choices instead of just one leads to an exponential improvement in the maximum load. This result demonstrates the power of two choices, and it has several applications to load balancing in distributed systems. In this thesis, we expand upon this result by examining related models and by developing techniques for stu...

The scenario is that a bulk data transfer is being performed over a TCP connection, from a host on a local area network (LAN) to a mobile host attached to the LAN by a radio link. In earlier work [10] we had assumed that packet losses in a TCP connection over a radio link are statistically independent. In this paper, we extend this analysis to a Rayleigh fading link, which we model by a two state Markov model. The bulk throughputs of TCP-OldTahoe and TCP-Tahoe are compared with and without fading, for various average signal-to-noise ratios. We also study the performance with a link protocol on the wireless link, and study the effect of varying the link packet size, the number of link packet attempts, and the vehicle speed. For the parameters of the BSD UNIX implementation, over a 1.5Mbps wireless link, we find that, with fading, a signal-to-noise ratio of at least 30dB is required to get reasonable throughput with TCP Tahoe or OldTahoe; this corresponds to at least 100 ti...

We propose a class of non-work-conserving service disciplines, called the Rate-Controlled Service Disciplines. When coupled with suitable admission control algorithms, Rate-Controlled Service Disciplines can provide end-to-end deterministic and statistical performance guarantees on a per-connection basis in an arbitrary topology packet-switching network. The key feature of a rate-controlled service discipline is the separation of the server into two components: a rate-controller and a scheduler. This separation makes it possible to obtain end-to-end performance characteristics by applying single node analysis at each switch. It also has several other distinct advantages: it decouples the allocation of bandwidths and delay bounds, uniformly distributes the allocation of buffer space inside the network to prevent packet loss, and allows arbitrary combinations of rate-control policies and packet scheduling policies. Rate-controlled service disciplines provide a general framework w...

The Shortest-Remaining-Processing-Time (SRPT) scheduling policy has long been known to be optimal for minimizing mean response time (sojourn time). Despite this fact, SRPT scheduling is rarely used in practice. It is believed that the performance improvements of SRPT over other scheduling policies stem from the fact that SRPT unfairly penalizes the large jobs in order to help the small jobs. This belief has led people to instead adopt “fair ” scheduling policies such as Processor-Sharing (PS), which produces the same expected slowdown for jobs of all sizes. This paper investigates formally the problem of unfairness in SRPT scheduling as compared with PS scheduling. The analysis assumes an M/G/1 model, and emphasizes job size distributions with a heavy-tailed property, as are characteristic of empirical workloads. The analysis shows that the degree of unfairness under SRPT is surprisingly small. The M/G/1/SRPT and M/G/1/PS queues are also analyzed under overload and closed-form expressions for mean response time as a function of job size are proved in this setting.

network model for studying the way throughput scales in a wireless network when the nodes are fixed, and showed that the throughput per source-destination pair is. Grossglauser and Tse (2001) showed that when nodes are mobile it is possible to have a constant or throughput scaling per source-destination pair. The focus of this paper is on characterizing the delay and determining the throughput-delay trade-off in such fixed and mobile ad hoc networks. For the Gupta-Kumar fixed network model, we show that the optimal throughput-delay trade-off is given by, where and are the throughput and delay respectively. For the Grossglauser-Tse mobile network model, we show that the delay scales as, where,.# is the velocity of the mobile nodes. We then describe a scheme that achieves the optimal order of delay for any given throughput. The scheme varies (i) the number of hops, (ii) the transmission range and (iii) the degree of node mobility to achieve the optimal throughput-delay trade-off. The scheme produces a range of models that capture the Gupta-Kumar model at one extreme and the Grossglauser-Tse model at the other. In the course of our work, we recover previous results of Gupta and Kumar, and Grossglauser and Tse using simpler techniques, which might be of a separate interest. Keywords: Stochastic processes/Queueing theory, Combinatorics, Information theory, Statistics.

The algorithms described in this thesis are designed to schedule cells in a very high-speed, parallel, input-queued crossbar switch. We present several novel scheduling algorithms that we have devised, each aims to match the set of inputs of an input-queued switch to the set of outputs more efficiently, fairly and quickly than existing techniques. In Chapter 2 we present the simplest and fastest of these algorithms: SLIP --- a parallel algorithm that uses rotating priority ("round-robin") arbitration. SLIP is simple: it is readily implemented in hardware and can operate at high speed. SLIP has high performance: for uniform i.i.d. Bernoulli arrivals, SLIP is stable for any admissible load, because the arbiters tend to desynchronize. We present analytical results to model this behavior. However, SLIP is not always stable and is not always monotonic: adding more traffic can actually make the algorithm operate more efficiently. We present an approximate analytical model of this behavior. SLIP prevents starvation: all contending inputs are eventually served. We present simulation results, indicating SLIP's performance. We argue that SLIP can be readily implemented for a 32x32 switch on a single chip. In Chapter 3 we present i-SLIP, an iterative algorithm that improves upon SLIP by converging on a maximal size match. The performance of i-SLIP improves with up to log 2 N iterations. We show that although it has a longer running time than SLIP, an i-SLIP scheduler is little more complex to implement. In Chapter 4 we describe maximum or maximal weight matching algorithms based on the occupancy of queues, or waiting times of cells. These algorithms are stabl...

We investigate the problem of providing a fair bandwidth allocation to each of n ows that share the outgoing link of a congested router. The buer at the outgoing link is a simple FIFO, shared by packets belonging to the n ows. We devise a simple packet dropping scheme, called CHOKe, that discriminates against the ows which submit more packets/sec than is allowed by their fair share. By doing this, the scheme aims to approximate the fair queueing policy. Since it is stateless and easy to implement, CHOKe controls unresponsive or misbehaving ows with a minimum overhead. 1 Introduction The Internet provides a connectionless, best eort, end-to-end packet service using the IP protocol. It depends on congestion avoidance mechanisms implemented in the transport layer protocols, like TCP, to provide good service under heavy load. However, a lot of TCP implementations do not include the congestion avoidance mechanism either deliberately or by accident. Moreover, there are a growi...

. We consider a distributed server system model and ask which policy should be used for assigning tasks to hosts. In our model each host processes tasks in First-Come-First-Serve order and the task's service demand is known in advance. We consider four task assignment policies commonly proposed for such distributed server systems: RoundRobin, Random, Size-Based, in which all tasks within a give size range are assigned to a particular host, and Dynamic-Least-Work-Remaining, in which a task is assigned to the host with the least outstanding work. Our goal is to understand the influence of task size variability on the decision of which task assignment policy is best. We find that no one of the above task assignment policies is best and that the answer depends critically on the variability in the task size distribution. In particular we find that when the task sizes are not highly variable, the Dynamic policy is preferable. However when task sizes show the degree of variability more charac...

Unknown references, also known as cold-start misses, arise during trace-driven simulation of uniprocessor caches because of the unknown initial conditions. Accurately estimating the miss ratio of unknown references, denoted by ¯, is particularly important when simulating large caches with short trace samples, since many references may be unknown. In this paper we make three contributions regarding ¯. First, we provide empirical evidence that ¯ is much larger than the overall miss ratio (e.g., 0.40 vs. 0.02). Prior work suggests that they should be the same. Second, we develop a model that explains our empirical results for long trace samples. In our model, each block frame is either live, if its next reference will hit, or dead, if its next reference will miss. We model each block frame as an alternating renewal process, and use the renewal-reward theorem to show that ¯ is simply the fraction of time block frames are dead. Finally, we extend the model to handle short trace samples an...