System Performance Modeling and Analysis of a Fault-Tolerant, Real Time Parallel Processor
Abstract
The use of fault-tolerant, real-time systems for the control of life- critical processes is becoming increasingly common, with examples including flight and nuclear reactor control systems. In such systems, the overhead associated with managing redundancy, communication, and task scheduling is critical due to real-time constraints imposed by the application; missed time deadlines can be viewed as system failures, with results as consequential as hardware failures. The Fault-Tolerant Parallel Processor (FTPP) was developed by Draper Laboratory as a fault-tolerant, real-time computing platform. This thesis analyzes the FTPP prototype operating system overhead through the use of empirical performance measurement and two performance models based on these measurements. One model is developed to predict the operating system overhead under various configurations and workloads; accurate prediction of overhead provides confidence that real-time constraints can be satisfied. Because the system communication overhead may vary depending upon the amount of contention by Processing Elements for service by the Network Element, a second model is developed to account for performance delays that may result from this contention. When such performance analysis and modeling are an integral part of a concurrent build-analyze-improve methodology, performance bottlenecks can be cost effectively removed at an early stage in development.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 27, 1993
- Accession Number
- ADA268070
Entities
People
- Robert J. Clasen
Organizations
- Air Force Institute of Technology