John M. McQuillan

The New Routing Algorithm for the ARPANET

The new ARPANET routing algorithm is an improvement over the old procedure in that it uses fewer network resources, operates on more realistic estimates of network conditions, reacts faster to important network changes, and does not suffer from long-term loops or oscillations. In the new procedure, each node in the network maintains a database describing the complete network topology and the delays on all lines, and uses the database describing the network to generate a tree representing the minimum delay paths from a given root node to every other network node. Because the traffic in the network can be quite variable, each node periodically measures the delays along its outgoing lines and forwards this information to all other nodes. The delay information propagates quickly through the network so that all nodes can update their databases and continue to route traffic in a consistent and efficient manner. An extensive series of tests were conducted on the ARPANET, showing that line overhead and CPU overhead are both less than two percent, most nodes learn of an update within 100 ms, and the algorithm detects congestion and routes packets around congested areas.

The ARPA network design decisions

Abstract A number of key decisions made in the design of the ARPA network over a five-year period serve as the context for an analysis of the fundamental properties and requirements of packet-switching networks and formulation of the fundamental criteria for evaluating network performance. The decisions described fall into the three major areas of network equipment design, store-and-forward subnetwork system design, and source-to-destination system design, and each decision is examined in detail.

A Review of the Development and Performance of the ARPANET Routing Algorithm

This paper presents a comprehensive review of the ARPANET routing algorithm, from its original implementation to our plans for future modifications. We hope that by collecting this information, and by providing considerable details, we can provide others with a useful reference document concerning some of the practical problems of network algorithm design. Much of the discussion below assumes a basic familiarity with the principles of packet switching, the ARPANET implementation, and some of the relevant terminology, information which can be found, for example, in [4]. Sections 1 and 2 give a brief summary of basic routing concepts and of the original routing algorithm, respectively. The following two sections describe in detail subsequent modifications and the actual implementation currently in use. Section 5 then discusses some problems that have developed over the past few years, as network usage has grown considerably. The final sections outline some explanations for these problems and some mechanisms for improving performance. We are in the process of implementing these and other changes to the routing algorithm.

An overview of the new routing algorithm for the ARPANET

The original routing algorithm of the ARPANET, in service for over a decade, has recently been removed from the ARPANET and replaced with a new and different algorithm. Although the new algorithm, like the old, is a distributed, adaptive routing algorithm, it is not similar to the old in any other important respect. In the new algorithm, each node maintains a data base describing the delay on each network line. A shortest-path computation is run in each node which explicitly computes the minimum-delay paths (based on the delay entries in the data base) from that node to all other nodes in the network. The average delay on each network line is measured periodically by the nodes attached to the lines. These measured delays are broadcast to all network nodes, so that all nodes use the same data base for performing their shortest-path computations. The new routing algorithm was extensively tested on the ARPANET before being released. This paper describes the algorithm and summarizes the results of these tests.

Issues in packet switching network design

The goals of this paper are to identify several of the key design choices that must be made in specifying a packetswitching network and to provide some insight in each area. Through our involvement in the design, evolution, and operation of the ARPA Network over the last five years (and our consulting in the design of several other networks), we have learned to appreciate both the opportunities and the hazards of this new technical domain.

Graph theory applied to optimal connectivity in computer networks

This is a report on some of the research that has been carried out in applying graph theoretical results to communications networks. The networks that we wish to investigate here consist of computers at various sites which are linked together by telecommunications circuits. Many of the results of graph theory may be applied to such networks; we will restrict ourselves to the consideration of connectivity. In designing a computer communications network, it is desirable to provide good connectivity among all sites at reasonable cost. For this reason, extremes like the fully-connected network (too expensive) and the star network (only as reliable as its center) are usually not considered. The connectivity constraints or reliability measures can be stated in different ways, and analytic techniques have been developed for some of these measures. Further, procedures for the synthesis of well-connected networks have also been invented. The reliability of communications networks is an important issue in their design and operation. Telecommunications circuits become noisy and unusable, and the communications computers may also fail. This report is a survey of that part of applied graph theory which is useful in the study of the connectivity of communications networks.

Alternatives for Data Network Architectures

The increasing use of computer data communications over the past several years has spawned a variety of network architectures to support requirements for distributed processing. Developed by various R&D groups,1-3by the common carriers,4-4by minicomputer and mainframe manufacturers,7,8and by the vendors of traditional communications hardware,9,10these new architectures represent alternative means to similar ends. This article provides a framework for understanding existing and forthcoming systems, focusing particular attention on the impact of evolving requirements and technologies.

Some considerations for a high performance message-based interprocess communication system

We continue to be concerned with interprocess communications systems (such as those described in references 1, 2, and 3 and called “thin-wire” communications systems in reference 4) which are suitable for communication between processes that are not co-located in the same operating system but rather reside in different operating systems on different computers connected by a computer communications network. Further, the systems with which we are concerned are assumed to communicate using addressed messages (e.g., reference 5) which are multiplexed onto the logical communications channel between the source process and the destination process, rather than using such traditional methods as shared memory (an impossibility for distributed communicating processes) or dedicated physical communications channels between pairs of processes desiring to communicate (which is considered to be prohibitively expensive).

Reliability issues in the ARPA network

Since the inception of the ARPA Network1 in 1969, we have been part of the group responsible for the development of that networks communications subnet. This role has provided us with a unique opportunity for study of the problems of network reliability and the effects of attempted improvements, particularly in the context of rapid network growth. Our overall philosophy for this effort has been that the network should be fault-tolerant with respect to individual component errors, and that the IMPs themselves should be fault-tolerant with respect to local failures. Along with this concern, we feel that the program should provide as much diagnostic information as possible. Component failures are of several kinds: hardware or software; solid, intermittent, or one-time. As we will discuss in the following sections, our attention has shifted in the last few years from handling circuit errors and failures to handling more difficult problems in the IMPs themselves: first intermittent problems, and recently even solid failures of major components.

Designing electronic mail systems that people will use

Physical Messages. When many system designers say electronic mall, they mean a modem equivalent for message-switching. For these people, modern electronic mall systems simulate the message-switching systems of old, station delivery is the norm, and the system is responsible for holding messages until deliverable. These systems have the great advantage of being able to replace existing systems or run in parallel with current systems without impact on the u~r. These systems focus very heavily on the location of the stations, the location of storage, and the location of sites. This is primarily the point of view of physical message delivery.