Browse Books

Reviewer: Michael Wolfe

This book describes how one might design and use a connection machine. The term “connection machine” is as loosely defined as “von Neumann machine,” but the general idea is that a Connection Machine (CM) is a computer with lots and lots of small processors, each with private memory, and all connected to each other through a network. The network is itself autonomous, working somewhat like a postal service: a message, once sent, arrives at the addressee without further intervention of either party and without passing through intermediate processors. Since the address does not specify the network path, the CM can continue to work in spite of processor failures and can adapt to network congestion. The Introduction in Chapter I gives some amusing and valid reflections on current computer design. For instance, von Neumann computers with semiconductor memories are becoming dominated by the size of the memory, which is idle most of the time. In order to get more efficiency, a design like the CM, with a little processor for each little chunk of memory, is required. Chapter II talks about programming the CM using LISP extensions. Without some knowledge of LISP, this will be hard to follow. Since the CM is designed with artificial intelligence applications in mind, the choice of LISP as the language is expected. The main LISP extension is an “xector,” which is a vector of LISP constructs. Xectors can be converted to lists and vice-versa new operators apply to xectors. Chapter III was the most interesting section, but not because of anything to do with connection machines or artificial intelligence. Here Hillis talks about design considerations for a new computer, such as size of a processor, serial vs. parallel arithmetic unit, communications network topology (the “topology zoo”), SIMD vs. MIMD, and so on. This is a good introduction to many of the terms used by computer architects (bandwidth, scalability, etc.); unfortunately, Hillis does not go into very much detail. The connection machine prototype, the CM-1, is described in Chapter IV. The CM-1 is an SIMD parallel processor, with 16,384 bit serial processors, connected in a Boolean n-cube topology. Each processor has 4096 bits of local memory. The key component of CM-1 is a custom VLSI chip containing 16 processors and one network router; because it is SIMD, this chip needs only one control unit. The CM-1 processor array is controlled by a microcontroller, which accepts instructions from a conventional host computer and broadcasts nanoinstructions to the processors. Each nanoinstruction has three single-bit inputs (two from memory and one flag register) and two outputs (one to memory and one flag register), and can produce any of the 65536 Boolean functions of three inputs and two outputs. Conditional operations based on the setting of a flag register are also allowed. Hillis calls this the ultimate RISC, with only one instruction. The network routers are also described, but not in as much detail. Chapter V talks about data structures (pointers vs. arrays) and Chapter VI discusses storage allocation (processor allocation and garbage collection). These are focused at implementing LISP on the CM. Chapter VII is “filler” comparing algorithms on a CM to physical laws. The annotated bibliography is thorough and perhaps useful. The book is not a lot of things. It is not a textbook about artificial intelligence or parallel computers. It is not a definition of connection machines, nor a description of the CM-1. It does not describe the language used for the CM-1, nor algorithms appropriate for the CM-1, nor parallel computers in general. Performance numbers for the CM-1 are missing entirely. It is not a good introductory text for parallel computers nor computers for artificial intelligence. It will not enlighten novices in the area nor convince unbelievers. It can be thought of as an extended Scientific American paper for an audience that understands LISP. Some design decisions for the CM-1 deserve special attention, but get little. For instance, by definition a connection machine has programmable connections; however, the CM-1 includes two special hardware processor connections. Each CM-1 processor has North, East, West, and South (NEWS) connections and connections to n-cube neighbors. Use of these connections will eliminate much of the flexibility of the CM, which is contrary to the basis of the design of the CM, and deserves more attention than it received. In summary, the book is well structured, but it contains either too much or too little detail to be really interesting.