DESIGN OF THE OPERATING SYSTEM FOR THE PASM PARALLEL PROCESSING SYSTEM
Abstract
The design of PASMOS, a distributed operating system for the PASM parallel processing system is considered. PASM is a partitionable SIMD/MIMD multimicrocomputer system. PASMOS has a hierarchical structure and is distributed throughout the hardware components of PASM. The architecture of the PASM system is overviewed and existing multiprocessor operating systems are surveyed. Parallel algorithms which compose a scenario for image contour extraction are developed. The scenario is used to demonstrate the operating system facilities required to support SIMD/MIMD task execution. The distributed software components of PASMOS and the facilities which these components provide to enable the system to create virtual SIMD/MIMD machines are described. The PASM command language, which allows users to interact with the PASM system, is presented. When the maximum task execution time is known a priori, task scheduling for a partitionable parallel processing system, such as PASM, is a two-dimensional bin packing problem. By applying a power of two constraint (given by the system architecture) on the number of processors which can compose a virtual machine, a new worst case bound is derived for Baker and Schwarz's first-fit shelf algorithm. Requiring the user to specify the maximum allowable execution time of a task before the task can be scheduled is a major limitation. A class of multiple-queue task scheduling algorithms which do not have this limitation is developed and applied to the PASM system. To improve processor utilization, it is desirable to overlap the operation of the secondary storage with computations being performed by the processors. Due to the dynamically reconfigurable architecture of PASM, a task must be preloaded prior to the schedulers final selection of processors. Two schemes which allow the system to preload input data and programs into the primary memories for upcoming tasks are presented and compared. The multiple control units in PASM share a common secondary storage for programs. The optimal service rate for the common secondary storage is determined. In summary, primary issues addressed include parallel algorithm design; operating system structure; user interaction; task execution environment; task management, scheduling, and preloading; memory management for the control units; and system performance.
Degree
Ph.D.
Subject Area
Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.