Computational science promises to become a third branch of discovery alongside theoretical science and experimental science. Its foundations are beginning to appear, if primarily in the context of individual published examples of its utility (and one Nobel Prize, awarded to Kenneth Wilson) and in books like those in this review. This little library of supercomputing books measures and records the maturation of vector supercomputing as well as the emergence of parallel processing. These current reference books acknowledge a handful of precursor volumes, most of which are no longer in print: Fernbach [1], DeVreese [2], Numich [3], Karin and Parker Smith [4], and Schneck [5].

With the exception of Christopher Lazou’s book, all of the reviewed books and most of their precursors are editorial compilations, mostly of research papers presented at conferences or symposia. Each has a somewhat different focus or integrating theme and may attract a different readership. From one point of view, the evolution of these conference themes over five years from 1985 to 1990 is an interesting yardstick for measuring the growth of computational science as a methodology during its most rapid years of development. While it has been said that one cannot compare apples and oranges, when presented with a tempting basket of fruit, some comparisons can be made. The books cluster generally into five categories: history and introduction to the field, early conferences, expert workshops, performance evaluation, and artificial intelligence.

Lazou

Lazou’s book is unique not just because it is a monograph, but because it is also a history, primer, and textbook that has already become a classic. First published in hardback in 1986, it was revised in 1988 as a paperback and reprinted again in 1989. I hope that the author will continue to update and re-release the book; it is a useful introduction for the scientific computing specialist who is trying to sort out a dazzling array of new computer architectures and their applications. Following a brief introduction, Lazou covers supercomputing architecture, building blocks, enabling technologies, and history from 1970. He then launches into such practical issues as program optimization, program vectorization, and performance tuning. Chapter 11 presents 12 application examples ranging from oceanography to film production using fractals. The last three chapters deal with future developments and new architectures; an appendix presents the features of FORTRAN 8X in the vector supercomputing context.

Matsen and Tajima; Kartashev and Kartashev

These two books are proceedings of early conferences on supercomputing held at the University of Texas in March, 1985 and in Santa Clara, California in May, 1988. Matsen and Tajima is a conference proceedings of papers in camera-ready form; it consists of 26 papers covering a variety of topics. Several of these papers are historically important because the ideas they presented have had substantial impact: examples are the early paper on numerical cosmology by Joan Centrella and her colleagues, Geoffrey Fox’s paper on the Caltech hypercube, the paper by Karl-Heinz Winkler and his colleagues on numerical fluid dynamics simulation, a paper on the Gibbs Project by Kenneth Wilson, and Stephen Wolfram’s paper on SMP, a precursor of Mathematica.

The book by the Kartashevs is a well-edited selection of papers addressing three topics: (1) the evolution and implementation of supercomputing architectures, (2) supercomputer hardware and software design, and (3) supercomputer applications and performance evaluation. The book is nicely produced and will be an enduring reference work. Of the 12 lengthy multi-author chapters, the Kartashevs’ introductory chapter on the evolution of supercomputer architecture, chapter 2 on the design of the Columbia Homogeneous Parallel Processor, and chapter 3 on the design of the Unisys Integrated Scientific Processor are particularly valuable from a technological historical perspective.

The paper by Milt Halem and his colleagues at Goddard Space Flight Center covers an interesting range of applications for both pipelined and massively parallel styles of supercomputing. The counterpoint paper on hypercube applications from Goddard’s rival sister institution JPL/Caltech is also a high water mark.

Lichnewsky and Saguez; Kowalik; Carey

These three edited proceedings of small expert workshops sponsored by government research agencies form a distinct group. Lichnewsky and Saguez’s book is based on an  INRIA  conference in Toulouse, France in 1986. Kowalik’s book covers a NATO–sponsored advanced research workshop in Trondheim, Norway in 1989. A National Science Foundation–sponsored workshop in Dallas in 1988 is the basis for Carey’s book.

The INRIA conference was one of the first conferences to focus on parallel supercomputing; specifically, parallel architectures, detection and exploitation of concurrency in software, parallel processor performance evaluation, and numerical algorithms able to exploit the potential gain of parallel processors. Up to that time, both general conferences and expert workshops had concentrated on vector supercomputers and their application. Lichnewsky and Saguez’s book is a workshop proceedings of the 21 presentations, 12 of which concern parallel processing methodology, two concern vector supercomputing, and the rest cover algorithm and performance evaluation.

The NATO conference, which was heavily attended by supercomputing providers, was evenly divided between presentations on supercomputer centers and networks, and presentations on parallel and distributed computing. Supercomputing vendors and centers also made presentations, mostly on topics in hardware and software optimization.

The National Science Foundation–sponsored workshop was devoted exclusively to the methods, algorithms, and applications of parallel supercomputing. The theme of the workshop was, “Now that parallel processing is a practical reality, how can scientists and engineers employ these new resources to solve problems not solvable on earlier systems?” Most of the papers are practical how-to-solve stories with worked examples and performance data. Carey’s work is organized as a multi-author book rather than a conference proceedings. After two introductory chapters by computer architects, the remaining 16 chapters focus on new algorithms for parallel supercomputers.

Martin; van der Steen

Martin’s and van der Steen’s books are devoted to performance evaluation of both established vector and advanced parallel supercomputers. Martin treats performance factors, measurements and metrics, and methods and models. The book contains five chapters on performance factors, four on measurements and metrics, and five on methods, models, and research directions in supercomputer performance evaluation. The material adds up to a pragmatic, state-of-the-art handbook on the subject.

Van der Steen is devoted to strategies for exploiting, evaluating, and benchmarking advanced computer architectures. The 18 chapters by a variety of authors represent a far wider--in fact, international--range of opinions on performance measurement. The editor presents the book as an overview or collection of current thinking on the topic and that is exactly what it is. Although less practical than  Martin,  it is considerably more entertaining, thanks to several polemical chapters.

Hwang and DeGroot

This book is in a class by itself, treating applications of high-performance multiprocessors primarily to problems in artificial intelligence. Seven chapters are devoted to numerical applications and the machines and software technology that support them. The remaining nine focus on parallel processing hardware and software for the solution of artificial intelligence problems. Although clearly aimed at the AI researcher looking to novel architectures for more computing power, the book covers traditional supercomputer applications well.

Comparison

This little library of supercomputing books measures and records the maturation of vector supercomputing and software during 1985–1990 and the emergence of parallel processing technology from its research infancy and childhood into its practical adolescence. Which is the best book? It depends on your needs. Lazou is the best introductory text or primer, but Carey would be the best advanced text for a course on current issues in computational science and engineering, or even an applied mathematics course on supercomputer applications. As a valuable reference work of papers documenting a range of architectural options, the Kartashevs’ book is good. Martin provides the most exhaustive treatment of supercomputing performance available. Hwang and  DeGroot  is peerless as a reference for parallel processing applications in artificial intelligence; the numerical supercomputing chapters are good in both coverage and depth.

Or do you want to survey this technology by reading only three books? Read Lazou, the Kartashevs, and Carey. Are you interested in preserving the good ideas that either weren’t implemented or didn’t survive in the marketplace? Read Matsen and Tajima, followed by the Kartashevs.

From the perspective of a bibliophile, the best book pedagogically is clearly Lazou, and the best produced book is the Kartashevs’ handsome work. The most useful to the scientist or applied mathematician is a toss-up between Carey and Kowalik--with perhaps an edge to Carey, which reads like a book and not just a conference proceedings.