Roger Eggen

Introducing parallel processing at the undergraduate level

With multiprocessor computers becoming more readily available on college campuses and with the recent development of software tools to aid in the implementation of parallel algorithms, it is time for parallel processing to enter the undergraduate curriculum. In this paper we show how widely available tools combined with concepts learned in an operating systems course could be used to develop an upper level undergraduate course in parallel processing. A typical artificial intelligence search problem (N-queens) is implemented in C and exhibits almost linear speedups.

Pattern matching in fewer comparisons

The problem of finding the occurrences of a pattern in a body of text is discussed. Several algorithms exist to solve this problem. The authors provide an overview of the Boyer-Moore algorithm, including its performance characteristics. They explain the implementation of the Kurak algorithm and its performance characteristics and compare the two algorithms. The analysis shows the Kurak algorithm to be superior when counting comparisons and when timing the search. The Kurak algorithm requires preprocessing of the text and thus, for the initial search, more time is required than for the Boyer-Moore algorithm. However, after the text has been processed the Kurak algorithm always provides superior performance.<<ETX>>

Key addressable data storage

The AK5 search tree is a popular and eflcient data storage technique yielding near optimal retrieval times when doing retrieval by key comparisons. By considering the internal bit configuration and doing some minor bit manipulation, the key can uniquely determine storage location in a tree or a two dimensioned array without using key comparisons. when using the bit configuration of the key and using a similar tree structure, we can store and retrieve data faster than an AI4 tree. Data can be stored directly into dynamically allocated two dimensioned array or the tree using the identical technique. The performance surpasses A VL trees even when a tree structure is used since all key comparisons are avoided. The technique relies solely on the bit configuration of the key to determine its location in the tree or its location in a two dimensioned ragged edged array. The paper &scribes the technique used to store and retrieve keys without comparisons.

A parallel algorithm for Steiner systems

The construction of Steiner systems provides an interesting foundation for the study of parallelism. Basic search techniques, such as depth first or breadth first search, warrant continued study when parallelism is involved. Though the emphasis of this paper is the study and corresponding,results of implementing a parallel algorithm solving Steiner systems, Steiner systems are important in their own right and have been studied for significant periods of time. The following areas are influenced by Steiner systems: 1) Graph theory [1], 2) Applications in Geometry [2], 3) Balanced incomplete block designs and groups [3], 4) Group theory [6], 5) Applications in algebra and design theory [7], and 6) Applications in graph theory [8]. In addition, the recent article by Colbourn and Van Oorschot in the ACM Computing Surveys attests to the multitude of applications of combinatorial designs in Computer Science [11].

A message based operating system promoting parallel programming

A description is given of a high level of programming where a graphical representation of a program is provided which uses C macros as a lower-level implementation tool. The programmer is challenged with the task of constructing the graph and the contents of the modules. The system interprets the graph, yielding a parallel executable form of the graph targeted for a real machine. A set of commonly used modules is maintained by the system library and the programmer adds the necessary modules when solving a specific problem. To add a module the programmer specifies data ports and the modules purpose. This allows other programmers to use modules previously created, enhancing programmer efficiency. The proposed system allows the programmer to concentrate on potential parallelism at a high level using predefined user-specific modules. Natural data flow is used in program construction, which requires little programmer training.<<ETX>>

An inherently parlle large grained data flow environment

A parallel programming environment based on data flow is described. Programming in the environment involves use with an interactive graphic editor which facilitates the construction of a program graph consisting of modules, ports, paths and triggers. Parallelism is inherent since data presence allows many modules to execute concurrently. The graph is executed directly without transformation to traditional representations. The environment supports programming at a very high level as opposed to parallelism at the individual instruction level.

Development of a parallel algorithmic language (abstract only)

Rolla* - A new style of programming is needed to meet the increased demands of computational speed. A software system based on dataflow and designed to use parallelism in future architectures is envisioned. The language model is designed so that any algorithm can be easily communicated to both man and machine. The basic entity of the model is that of units which are interconnected by a series of data paths like pipes. A unit is any process, file or device capable of consuming or supplying data. Consuming and supplying pairs are categorized by the type of flow of data between units. A notation has been developed to express the dataflow graphs which exposes the parallel nature of the model. Ports, index numbers and process names allow an unambiguous representation of the graphs and support nesting to any level. Further research is being done in the area of unit description and dataflow control.