Abstract
Due to the inherent potential of parallel processing, a lot of attention has focused on massively parallel computer architecture. To a large extent, the performance of a massively parallel architecture is a function of the flexibility of its communication network. The ability to configure the topology of the machine determines the ease with which problems are mapped onto the architecture. If the machine is sufficiently flexible, the architecture can be configured to match the natural structure of a wide range of problems. There are essentially four unique types of massively parallel architectures: 1. Cellular Arrays 2. Lattice Architectures [21, 30] 3. Connection Architectures [19] 4. Honeycomb Architectures [24] All four architectures are classified as SIMD. Each, however, offers a slightly different solution to the mapping problem. The first three approaches are characterized by easily distinguishable processor, communication, and memory components. In contrast, the Honeycomb architecture contains multipurpose processing/communication/memory cells. Each cell can function as either a simple CPU, a memory cell, or an element of a communication bus. The conventional approach to massive parallelism is the cellular array. It typically consists of an array of processing elements arranged in a mesh pattern with hard wired connections between neighboring processors. Due to their fixed topology, cellular arrays impose severe limitations upon interprocessor communication. The lattice architecture is a somewhat more flexible approach to massive parallelism. It consists of a lattice of processing elements embedded in an array of simple switching elements. The switching elements form a programmable interconnection network. A lattice architecture can be configured in a number of different topologies, but it is still only a partial solution to the mapping problem. The connection architecture offers a comprehensive solution to the mapping problem. It consists of a cellular array integrated into a packet-switched communication network. The network provides transparent communication between all processing elements. Note that the communication network is physically abstracted from the processor array, allowing the processors to evolve independently of the network. The Honeycomb architecture offers a unique solution to the mapping problem. It consists of an array of identical processing/communication/memory cells. Each cell can function as either a processor cell, a communication cell, or a memory cell. Collections of Honeycomb cells can be grouped into multicell CPUs, multi-cell memories, or multi-cell CPU-memory systems. Multi-cell CPU-memory systems are hereafter referred to as processing clusters. The topology of the Honeycomb is determined at compilation time. During a preprocessing phase, the Honeycomb is adjusted to the desired topology. The Honeycomb cell is extremely simple, capable of only simple arithmetic and logic operations. The simplicity of the Honeycomb cell is the key to the Honeycomb concept. As indicated in [24], there are two main research avenues to pursue in furthering the Honeycomb concept: 1. Analyzing the design of a uniform Honeycomb cell 2. Mapping algorithms onto the Honeycomb architecture This technical report concentrates on the first issue. While alluded to throughout the report, the second issue is not addressed in any detail.
Date of this Version
12-1-1987
Comments
Project was supervised by Professor V. Milutinovic. His activities in relation to this project were sponsored by the NCR Corporation, World Headquarters, Dayton, Ohio