1. Introduction

Among the services to be offered by a broadband ISDN network, those entailing some sort of performance guarantee are particularly interesting, especially for networks based on the packet-switching principle. We call them real-time communication services. Since packet switching (or cell switching) seems to be gaining the upper hand in the argument about how B-ISDNs should operate, the question of whether and how performance can be guaranteed in such a congestion-prone environment as that created by packet switching is much more than an intriguing academic problem.

The Tenet Group at the University of California at Berkeley and the International Computer Science Institute gave this question an affirmative answer in the case of a network with arbitrary topology, consisting of hosts and homogeneous store-and-forward switches, in which packet transmissions are scheduled by a multi-class version [FeVe90a] of the Earliest Due Date (EDD) deadline-based discipline [LiLa73]. In the sequel, whenever we use the term simple network, we shall refer to such a network. The details of the scheme that applies to simple networks have been published in [FeVe90a], where a general method and the specific algorithms for establishing real-time connections are given, in [FeVe90b], where the allocation of buffer space to such connections is discussed, and in [VeZF91], where a scheme for distributed jitter control is presented. The Tenet approach allows a client of the real-time communication service to choose without restrictions the desired bounds of the important performance indices (delay, throughput, reliability) and to specify, also without restrictions, the values of the parameters of a simple traffic characterization. This information is used to determine whether the client's request can be accepted by the network in its current load conditions.

The Tenet approach differs in many respects from other solutions that provide performance guarantees in a packet-switching network. Unlike several other approaches, it is not based on the use of a particular type of switch or gateway, nor is it confined to only one kind of network topology. Unlike the Flow Protocol [Zhan90], which offers average throughput bounds and, as a secondary and constrained objective, delay bounds, it allows clients to specify and obtain throughput and delay bounds totally independently of each other, as well as jitter and loss bounds. The Session Reservation Protocol [AnHS90] is based on a philosophy similar to ours, but with several important differences; these include a different admission control policy, a different traffic model, the bundling of control and delivery functions within the same protocol, and the compatibility of the SRP header with the IP header, which raises some obstacles to internetworking. The Stream Protocol, Version II (ST-II) [Topo90] is an experimental internetwork-layer protocol developed for the DARPA Internet. Unlike our current scheme, it offers multicast connections; its traffic and performance parameters, and its channel establishment procedure, differ appreciably from those of our approach. No resource reservation policies are detailed in the specification of ST-II, though the usable policies are obviously constrained by the messages exchanged during channel establishment and by the parameters contained in those messages. Another protocol at the same level, the Multipoint Congramoriented High-performance Internet Protocol (MCHIP) [PaTu90], provides multicast capabilities, and may make use of resource servers, which keep track of the channels established and of the available resources. The establishment procedure and the resource reservation policies adopted for MCHIP have not, to our knowledge, been published yet. The Asynchronous Time Sharing (ATS) approach differs from ours primarily because of its being based on a fixed menu of quality-of-service classes [LaPa91]: each real-time