Workload, Performance, and Reliability of Digital Computing Systems.

TLDR: A new modeling methodology to characterize failure processes in Time-sharing systems due to hardware transients and software errors is summarized, which gives quantitative relationships between performance, workload, and (lack of) reliability for digital compiuting systems.

Abstract: In this paper a new modeling methodology to characterize failure processes in Time-sharing systems due to hardware transients and software errors is summarized. The basic assumption made is that the instantaneous failure rate of a system resource can be approximated by a deterministic function of time plus a zero-mean stationary Gaussian process, both depending on the usage of the resource considered. The probability density function of the time to failure obtained under this assumption has a decreasing hazard function, partially explaining why other decreasing hazard function densities such as the Weibull fit experimental data so well. %rttiermore, by considering the Kernel of the Operating System as a sysSem resource, this methodology sets the basis for independent methods of evaluating the contribution of software to system unreliability, and gives some non obvious hints about how system reliability could be improved. A real system has been characterized according to this methodology, and an extremely good fit between predicted and observed behavior has been found. Also, the predicted system behavior according to this methology is compared with the predictions of other models such as the exponential, Weibull. and periodic failure rate. The work presented in this paper describes a new modeling methodology. This methodology gives quantitative relationships between performance, workload, and (lack of) reliability for digital compiuting systems. Current methodologies for reliability assessment may provide good models for explaining and predicting the behavior of systems in the presence of hard (recurrent) faults, but the effect and charcterization of transient (non recurrent) faults and software (either design or implementation) errors is still. very elusive. These current reliability measures do not give individual users a feeling of the impact of unreliability on performance in genetpal purpose system operating under a variety of workloads. That is, there are no general methods for a quantitative assessment of the! benefits of fault.tolerance. 2 Prior work