The activity in building gigabit speed networks has led many researchers to re-examine the issue of congestion control. We describe a rate-based hop-by-hop congestion control mechanism in which the service rates of connections are dynamically adjusted at a switch, using feedback information provided by the neighboring switches. The desired service rate is computed based on a control equation that utilizes a model of the system with feedback information used to correct inaccuracies in the model. We use an analytical model to prove that the expected value of the queue occupancy and throughput of a controlled connection converge to the desired operating point. We also study the variation of the queue occupancy and throughput in steady-state as well as the transient response. The analytical results provide insights into how the parameter values chosen affect performance. We use simulations to compare the performance of the scheme with an equivalent end-to-end control scheme. Our analytical and simulation results show that the hop-by-hop scheme reacts faster to changes in the traffic intensity and, consequently, utilizes resources at the bottleneck better and loses fewer packets than the end-to-end scheme.

Future national and international telecommunication networks are headed towards using ATM (asynchronous transport multiplexing) technology. ATM technology allows seamless integration between different traffic types such as voice, video and data on a single network and also between customer premises' equipment and common carriers' networks. Unlike the current telephone networks, the ATM network provides virtual connections that have variable bandwidth and variable path delay between a sender and a receiver. Although ATM networks do support constant bandwidth channel as one of the traffic classes, using virtual connections as variable bandwidth pipes for video traffic is more attractive for reasons of efficiency in network usage and video quality. In this paper, we show that it is feasible to use congestion signals from networks as a feedback to the encoder and achieve very good perceptual quality for real-time video transmissions.

We show that modulating the source rate of a video encoder based on congestion signals from the network has two major benefits: the quality of the video transmission degrades gracefully when the network is congested and the transmission capacity is used efficiently. Source rate modulation techniques have been used in the past in designing fixed rate video encoders used over telephone networks. In such constant bit rate encoders, the source rate modulation is done using feedback information about the occupancy of a local buffer. Thus, the feedback information is available instantaneously to the encoder. In the scheme proposed, the feedback may be delayed by several frames because it comes from an intermediate switching node of a packet switched network. The paper shows the proposed scheme performs quite well despite this delay in feedback. We believe the use of such schemes will simplify the architecture used for supporting real time video services in future nationwide gigabit networks.

In this paper we show that modulating the source rate of a video encoder based on feedback information from the network results in graceful degradation in picture quality during periods of congestion. Such source rate modulation techniques have been used in the past in designing video encoders used to generate data at a fixed rate. In such constant bit rate encoders, the source rate modulation is done using feedback information about the occupancy of a local buffer. Since the buffer is local, the feedback information is available instantaneously to the encoder. In the proposed scheme, the feedback information is delayed because it comes from within a packet-switching network. The feedback provides information about the traffic at switches along the path of the video connection. We show that the proposed scheme performs well, even though the information is delayed for relatively large intervals of time. We believe that the use of such schemes will simplify the architecture used for supporting real-time services in future nationwide gigabit networks.

The flow/congestion control scheme of TCP is based on the sliding window mechanism. As we demonstrate in this paper, the performance of this and other similar end-to-end flow control schemes deteriorates as networks move to the gigabit range. This has been the motivation for our search for a new flow and congestion control scheme. In this paper, we propose as an alternative, a hop-by-hop rate-based mechanism for congestion control. Due to the increasing sophistication in switch architectures, to provide “quality of service” guarantees for real-time as well as bursty data traffic, the implementation of hop-by-hop controls has become relatively inexpensive. A cost-effective implementation of the proposed scheme for a multi-gigabit packet switch is described in [2]. In this paper, we present results of a simulation study comparing the performance of this hop-by-hop flow control scheme to two end-to-end flow control schemes. The results indicate that the proposed scheme displays stable behavior for a wide range of traffic conditions and diverse network topologies. More importantly, the performance of the scheme, measured in terms of the average number of occupied buffers, the end-to-end throughput, the network delay, and the link utilization at the bottleneck, is better than that of the end-to-end control schemes studied here. These results present a convincing case against popular myths about hop-by-hop control mechanisms.

Current host network interface protocols limit performance visible to applications, waste critical host resources such as system bus and main memory bandwidth, and offer little protection against network malfunctions or hostile remote users. The authors present UNDIP, a host-network interface protocol that addresses these problems. The protocol is adaptable to networks with widely different switching techniques, different host system architectures, and different transport protocols. The authors describe the key design aspects as well as an implementation of UNDIP on a prototype network adapter board for the VMP multiprocessor system. They also show the performance advantage of UNDIP using preliminary measurements derived from the prototype.>

High performance computer communication between multiprocessor nodes requires significant improvements over conventional host-to-network adapters. Current host-to-network adapter interfaces impose excessive processing, system bus and interrupt overhead on a multiprocessor host. Current network adapters are either limited in function, wasting key host resources such as the system bus and the processors, or else intelligent but too slow, because of complex transport protocols and because of an inadequate internal memory architecture. Conventional transport protocols are too complex for hardware implementation and too slow without it.

We consider two styles of executing a single job or an algorithm: either the job is subdivided into tasks, each of which is executed. on a separate processor, or the entire job is executed on a single processor, that has the same capacity as the sum of the processors in the earlier case. The algorithm is abstracted as consisting of a number of tasks with dependencies among them. Our model of dependencies among tasks allows sequential execution, parallel execution, synchronization, and spawning of tasks. The model assumes that the dependencies are known before the job begins, and a task in not preempted after its execution begins. With the usual assumptions such as exponential distribution of task execution times, and Poisson arrival of input data, we are able to show that the centralized execution completes the job faster than the decentralized execution only for a certain range of parameters of algorithms. We also give counterexamples that show that, contrary to popular belief, the reverse is true for some values of parameters of algorithms.