In this paper, we share our experience in designing and building a content-based switch which we call L5. In addition to the layer 2-3-4 information available in the packet, a content-based switch uses application level information to route traffic in the network. Making routing decisions based on information contained in the payload is not a new idea. In fact application level proxies which are functionally equivalent to a content-based switch, have been around for years. Our contribution is in combining the functionalities of an application level proxy with the data handling capabilities of a switch into a single system. In this paper, we describe the architecture of the L5 system along with the details of how application level information can be efficiently processed in the switch hardware. We cover two specific application examples that we believe are ideal candidates for content-based switching: one is routing HTTP sessions based on uniform resource locators (URL) and the other is session-aware dispatching of secure socket layer (SSL) connections.

The last couple of years has seen a growing momentum towards using the Internet for conducting business. One of the key enablers for business applications is the ability to setup secure channels across the Internet. The Secure Sockets Layer (SSL) protocol provides this capability and it is the most widely used transport layer security protocol. In this paper we investigate the performance of SSL both from a latency as well as a throughput point of view. Since SSL is primarily used to secure Web transactions, we use the SPECWeb96 benchmark suitably modified for use with the SSL protocol. We benchmark two of the more popular Web servers that are in use today and find that they are a couple of orders of magnitude slower when it comes to serving secure Web pages. We investigate the reason for this deficiency by instrumenting the SSL protocol stack with a detailed profiling of the protocol processing components. Based on our findings we suggest two modifications to the protocol that reduce the latency as well as increase the throughput at the server.

This paper investigates the problem of making QoS guarantees available in access devices such as edge routers, that are commonly deployed in today's IP networks. We propose a specific design which we evaluate by carrying out a complete implementation, whose performance we then measure in the context of an experimental testbed. The results indicate that a reasonable level of service differentiation, i.e., rate and delay guarantees, can be provided with a minimal impact on the raw packet forwarding performance of edge devices.

In this paper, we review the basic mechanisms used in packet networks to support Quality-of-Service (QoS) guarantees. We outline the various approaches that have been proposed, and discuss some of the trade-offs they involve. Specifically, the paper starts by introducing the different scheduling and buffer management mechanisms that can be used to provide service differentiation in packet networks. The aim is not to provide an exhaustive review of existing mechanisms, but instead to give the reader a perspective on the range of options available and the associated trade-off between performance, functionality, and complexity. This is then followed by a discussion on the use of such mechanisms to provide specific end-to-end performance guarantees. The emphasis of this second part is on the need for adapting mechanisms to the different environments where they are to be deployed. In particular, fine grain buffer management and scheduling mechanisms may be neither necessary nor cost effective in high speed backbones, where “aggregate” solutions are more appropriate. The paper discusses issues and possible approaches to allow coexistence of different mechanisms in delivering end-to-end guarantees.

In recent years, a number of link scheduling algorithms have been proposed that greatly improve upon traditional FIFO scheduling in being able to assure rate and delay bounds for individual sessions. However, they cannot be easily deployed in a backbone environment with thousands of sessions, as their complexity increases with the number of sessions. In this paper, we propose and analyze an approach that uses a simple buffer management scheme to provide rate guarantees to individual flows (or to a set of flows) multiplexed into a common FIFO queue. We establish the buffer allocation requirements to achieve these rate guarantees and study the trade-off between the achievable link utilization and the buffer size required with the proposed scheme. The aspect of fair access to excess bandwidth is also addressed, and its mapping onto a buffer allocation rule is investigated. Numerical examples are provided that illustrate the performance of the proposed schemes. Finally, a scalable architecture for QoS provisioning is presented that integrates the proposed buffer management scheme with WFQ scheduling that uses a small number of queues.

An important challenge for the future growth of the Internet is to design routers that can forward the exponentially increasing volume of traffic, and at the same time provide the service differentiation needed by new applications. In this paper, we describe the architecture, implementation, and initial experiences with a system designed to meet this challenge. This system, which we call a QoS capable switch-router (QSR), combines the salient features of switching and routing technologies to provide high throughput and support the different classes of service being defined by the IETF. It consists of a core (ATM) switch fabric connecting intelligent adapters, each capable of both routing and switching pockets. A control engine is responsible for routing, RSVP signalling, and resource management. We have built a prototype network of 3 systems connected to several UNIX hosts, and have conducted preliminary performance measurements on this network.

This paper addresses the problem of providing per-connection end-to-end delay guarantees in a high-speed network. We assume that the network is connection oriented and enforces some admission control which ensures that the source traffic conforms to specified traffic characteristics. We concentrate on the class of rate-controlled service (RCS) disciplines, in which traffic from each connection is reshaped at every hop, and develop end-to-end delay bounds for the general case where different reshapers are used at each hop. In addition, we establish that these bounds can also be achieved when the shapers at each hop have the same "minimal" envelope. The main disadvantage of this class of service disciplines is that the end-to-end delay guarantees are obtained as the sum of the worst case delays at each node, but we show that this problem can be alleviated through "proper" reshaping of the traffic. We illustrate the impact of this reshaping by demonstrating its use in designing RCS disciplines that outperform generalized processor sharing-based service disciplines.

In this paper, we investigate some issues related to the efficient provision of end-to-end delay guarantees in the context of the Guaranteed (G) Services framework [16]. First, we consider the impact of reshaping traffic within the network on the end-to-end delay, the end-to-end jitter, as well as per-hop buffer requirements. This leads us to examine a class of traffic disciplines that use reshaping at each hop, namely rate-controlled disciplines. In this case, it is known that it is advantageous to use the Earliest Deadline First (EDF) scheduling policy at the link scheduler [8]. For this service discipline, we determine the appropriate values of the parameters that have to be exported, as specified in [16]. Subsequently, with the help of an example, we illustrate how the G service traffic will typically underutilize the network, regardless of the scheduling policy used. We then define a Guaranteed Rate (GR) service, that is synergetic with the G service framework and makes use of this unutilized bandwidth to provide rate guarantees to flows. We outline some of the details of the GR service and explain how it can be supported in conjunction with the G service in an efficient manner.

This paper addresses the problem of providing per-connection end-to-end delay guarantees in a high-speed network. We consider a network comprised of store-and-forward packet switches, in which a packet scheduler is available at each output link. We assume that the network is connection oriented and enforces some admission control which ensures that the source traffic conforms to specified traffic characteristics. We concentrate on the class of rate-controlled service (RCS) disciplines, in which traffic from each connection is reshaped at every hop, and develop end-to-end delay bounds for the general case where different reshapers are used at each hop. In addition, we establish that these bounds can also be achieved when the shapers at each hop have the same "minimal" envelope. The main disadvantage of this class of service discipline is that the end-to-end delay guarantees are obtained as the sum of the worst-case delays at each node, but we show that this problem can be alleviated through "proper" reshaping of the traffic. We illustrate the impact of this reshaping by demonstrating its use in designing RCS disciplines that outperform service disciplines that are based on generalized processor sharing (GPS). Furthermore, we show that we can restrict the space of "good" shapers to a family which is characterized by only one parameter. We also describe extensions to the service discipline that make it work conserving and as a result reduce the average end-to-end delays.

We consider an important tradeoff between processor and memory allocation in distributed parallel processing systems. To study this tradeoff, we formulate stochastic models of parallel program behavior, distributed parallel processing environments and memory overheads incurred by parallel programs as a function of their processor allocation. A mathematical analysis of the models is developed, which includes the effects of contention for shared resources caused by paging activity. We conduct a detailed analysis of real large-scale scientific applications and use these results to parameterize our models. Our results show that memory overhead resulting from processor allocation decisions can have a significant effect on system performance in distributed parallel environments, strongly suggesting that memory considerations must be incorporated in the resource allocation policies for parallel systems. We also demonstrate the importance of the inter-locality miss ratio, which is introduced in this paper and analyzed for the first time.