A study of the reliability of token rings designed according to the IEEE 802.5 standard is described. The reliability model is based on information in the standards, plus implementation details, and is more realistic than previous models. Although the analysis is approximate, it includes all factors thought to be important in determining reliability, since an approximate analysis which included all such factors was felt to be more useful than an exact analysis which ignored important factors. The precise reliability measure analyzed is availability of a connection from a specified source to a specified destination. Although the use of an active communications medium instead of a passive bus may decrease reliability of the token ring, the architecture gives great flexibility in bypassing or otherwise recovering from faults. For large networks this adds considerably to the reliability. If many devices are inactive, however, this advantage can be eliminated by even a small probability that inactive devices are not successfully bypassed.>

FISHNET (facilities integrated in a shared habitat network) addresses the problem of effectively integrating a wide variety of computers, terminals, memory devices and computer peripherals in a local environment. High performance is achieved by effectively separating the high volume node-to-node data communications and the low-volume network control information. Data messages travel through point-to-point data links. Control messages, which set up data paths and coordinate network activities, use an independent multidrop control link which connects all devices. The fully distributed control mechanism allows a high degree of concurrent data transmission.>

This paper is a critique of major weaknesses of current techniques for analyzing the reliability of computer communication networks. Several of the most important factors influencing the reliability seen by users of such systems have been almost completely ignored or modeled in a rudimentary fashion. Ignoring such factors may be partially justified by the inherent complexity of the problems being studied, which often precludes exact computations being feasible for networks of any size, even with simplifying assumptions. Nevertheless, the results of models which ignore such factors are questionable even if they are computable. The focus of models of such networks should be on the effects of system malfunctions, degraded performance, excessive delays, etc., on system users. New approximate modeling techniques appear to be the most realistic approach to studying such problems.

Published data describing the probability distributions of data communication system reliability parameters such as availability are very sparse. This paper describes a set of data collected in order to model communication line availability by monitoring operational communication systems. Probability distribution functions for single line availabilities and for the durations of failures are given. A heuristic model for estimating availabilities for more complex systems is presented along with computer simulation data validating the accuracy of the model. The results presented make possible more realistic availability predictions for common types of networks than have previously been computable.

This paper presents a new modular loop/ring architecture combining advantages of several earlier centralized and decentralized ring-structured loop networks while remaining simple. This is accomplished by introducing two major innovations: first, use of a separate control loop for control messages flowing between nodes and a loop controller; second, dynamically implementing partitionable segments between adjacent nodes and treating them as separate links for data transmission.

This paper presents a simplified approach to developing performance models for complex systems, such as distributed processing systems, and illustrates the approach by applying it to studying the performance of a new ring-structured microcomputer network. The approach utilizes both regression (curve-fitting) techniques and queueing theory analysis, and suggests that response times of the system studied can be closely approximated by a multiserver queueing model which is generalized to allow a noninteger number of servers. Both the approach and the approximate model developed appear to be applicable to a wide variety of systems.

Although a great deal of work on modeling reliability of data communications systems has been done, there are important problems with current models which do not seem to be adequately appreciated, much less solved. Two important problems of this nature are discussed: 1) finding adequate techniques for computing reliabilities of systems with dependent communications line failures (which are the dominant cause of system failures in important classes of systems), and 2) developing adequate techniques for handling systems with tremendously variable reliability parameters (since parameters such as line failure rates typically vary by three or four orders of magnitude for different communication lines). Motivations for considering these to be important problems are given and the results of initial rough models which include the effects are discussed.

Some queueing theory formulas for loop systems with random slot seizure previously obtained by the author via a lengthy series of derivations are shown to follow immediately from priority queueing formulas derived by Takács if an appropriate artifice is used to make Takács' results applicable.

Although almost all published analyses of the reliability of communication systems assume failures of different lines are independent, such failures are actually highly correlated. A description of why this is true is given. This is followed by an approximate theoretical analysis of some measurement data. Both approaches yield the inescapable conclusion that assuming failures of different communications lines serving one location are independent can lead to dangerously inaccurate reliability predictions. The significance of these results for the designer of highly reliable systems involving data communications is discussed.

A simple analytical model is developed for loops with centralized control, random slot seizure, and fixed slot duration. Slots are assumed to be generated at regular intervals, regardless of whether they are utilized or not, while requests for service at terminals attached to the loop are described by Poisson arrival statistics, For loops used for data collection, mean waiting-line lengths and queue lengths are obtained both at instants when slots are due to arrive and at instants of arrivals of requests for service. Mean waiting times and queueing times are computed and the relative efficiencies of different loop sizes compared. Simultaneous data collection and distribution is also considered and the degradation in data-collection service due to use of resources for data distribution is plotted.

The manner in which the performance of a communication processor (CP) can be expected to deteriorate under extreme loading conditions has been studied through general-purpose simulation system (GPSS). Processing of characters to determine their disposition when they initially arrived at the CP was assumed to be the primary factor limiting performance under extreme loads and was hence modeled carefully. The CP proved to have an extremely sharp saturation threshold, with the threshold very sensitive to minor variations in character processing times. Simple design criteria for ensuring satisfactory performance without resorting to an extreme worst case design were developed. Also, the effects of some changes in buffering and in the use of priorities were determined.

This is a review paper discussing and comparing most of the approaches to learning without a teacher which have been suggested to date. The approaches discussed are divided into six classes: guessing a sequence of hypotheses, modifications of this first approach, approximating probability densities by others more easily computed, estimating parameters of a known decision rule, theoretically exact methods which require approximations in implementation, and some miscellaneous approaches. At present, all these approaches are known to be feasible, but few useful methods of comparing them to see which are best have been developed.

Alterations in probability densities produced by iterative application of Bayes' rule are analyzed. Computational requirements for finding a sequence of a posteriori densities remain reasonable despite a growing sample size if and only if a sufficient statistic expressible as a vector of fixed dimension exists. The existence of such a sufficient statistic also insures existence of reproducing a priori probability densities (a priori probability densities insuring that the a posteriori densities are in the same family). The theory is applied to find a class of sequential Bayes estimators for a Gaussian covariance matrix, and to treat a variety of adaptive "Bayesian learning" schemes in a unified manner.