Armstrong, A loosely-coupled multiprocessor testbed for reconfigurable topologies

Abstract

Armstrong is a loosely-coupled parallel-processing system designed and constructed in the Laboratory for Engineering Man/Machine Systems (LEMS) in the Brown University Division of Engineering. The system was built in order to provide a useful, controllable, research-oriented parallel-processing testbed for this dissertation. The main purpose of the dissertation (in addition to building the hardware) was to develop an appropriate distributed operating system for this architecture by examining real applications under complete monitoring and network control. The Armstrong hardware consists of up to 100 identical processing nodes, each containing a Motorola 68010 microprocessor, a National 32081 floating-point coprocessor, 512 kilobytes of memory, virtual memory support, and eight point-to-point links operating at 40 megabits per second. There is no shared memory in the system. The operating system supports multitasking on each node, and message-passing between processes on the same node or different nodes. The message-passing software provides idealized functionality, including arbitrary-length messages, complete error checking and recovery, and flow control. There are several research contributions in this dissertation. First, the communication network is manually reconfigurable into any arbitrary graph with a degree of eight or less. In addition, the operating system supports this reconfigurability by providing location-independent communication and automatic shortest-path message routing. User programs are not aware of the current network topology. Thus, applications can be benchmarked on different networks without even recompiling them. Finally, a number of applications have been implemented and benchmarked on Armstrong, including Mandelbrot fractal computation, two-dimensional discrete Fourier transforms, simulated annealing, real-time spectrogram generation, and robot control. A discussion of these applications and an analysis of their performance on the actual hardware is presented.