Abstracts computing research laboratory
SIGACT News Staff · ACM SIGACT News · 1983
Performance measures used in contemporary analyses suffer from a number of shortcomings when real-time multiprocessors are considered.Most of these measures do not take account of the needs of the application, except perhaps in a sub-Jective manner.In this pap.er,we consider some new performance measures to characterize fault-tolerant multiprocessors used in the control of critical processes.Our performance indices are based on controller response time.By relating this to the needs to the application, we have been able to derive indices that faithfully reflect the performance of the multiprocessor in the context of the appticatio~, that permit the objective comparison of rival computer systems, and that can either be definitively estimated or objectively measured.An example of a controller in an idealized satellite application is provided. CRL-TR-2-82, Kang G. Shin and Yann-Hang Lee "Analysis of the Impact of Error Detection on Computer Performance,"abstract -Conventionally, reliability analyses either assume that a fault/error is detected immediately following its occurrence, or neglect damages caused by latent errors.Though unrealistic, this assumption has been imposed in order to avoid the difficulty of determining the respective probabilities that a fault induces an error and the error is then detected in a random amount of time after its occurrence.As a remedy for this problem, in this paper a model is proposed to analyze the impact of error detection on computer performance under moderate assumptions.Error latency -the time interval between occurrence of an error and the moment of error detection -is used to measure the effectiveness of a detection mechanism.We have used this model to ( 1) predict the probability of producing an unreliable result, and (2) estimate the loss of computation due to fault and/or error.94 CRL-TR-3-82.B.A. Makrucki and T.N.Mudge "A Stochastic Model of Parallel and Concurrent Program Execution on Multiprocessors," abstract -This report summarizes a model developed to allow the evaluation of parallel program execution on multiprocessors.The model is intended for MIMD algorithms in which the individual processors are coupled through their programs' interaction with memory.The model is not intended for SIMD algorithms.Specifically, estimates of processor utilization, execution times of programs or subprograms, and memory bandwidth can be obtained from the model.Earlier research has concentrated on the last of these quantities and a body of research, which might be termed "memory interference models", has evolved.The work ,eported here goes one step further by allowing the programs to be included in the model.Consequently questions about the performance of the processors running the p~'og~ams c~n also be answered. CRL-TR-4-82.