A PREDICTIVE MODEL FOR THE CAPACITY MANAGEMENT OF UNIX SYSTEMS

Abstract

Traditional performance evaluation and capacity planning carried out during performance studies of a computer system require a lot of effort on the part of the performance analyzer. In addition, these studies typically necessitate the collection of huge amounts of performance data at those periods when the computer system is experiencing performance bottlenecks. This tends to have an aggravating effect on the performance of an already overloaded system, which can be a rather undesirable situation if these are peak periods when providing an acceptable service level to users is of utmost importance. The amount of data which has to be collected may also result in the information overload problem for the performance analyzer. To overcome the difficulties discussed above, a performance model based on predicting the resource usage of applications is proposed. The model attempts to predict potential problems by taking an applications point of view as opposed to a systems viewpoint taken in traditional performance measurement. The model addresses the above difficulties by collecting the relatively "static" (across computer installations) performance data previously collected during periods of poor performance as early as possible, before the system is in operation or encounters bottlenecks. The collected data are then used to predict the resource usage of the applications comprising the potential workload of a system, given the forecasted. workload characteristics. The “static" performance data comprises of the resource consumption of components of the applications, right down to the resource usage of the basic building blocks of the programs used in the applications. From the predicted resource usage of each application in the potential workload, it is then not difficult to identify applications which are likely to compete for certain system resources when they are being executed concurrently in the system. The competition for scarce system resources is the main cause of performance problems. In this study, a 486DX Personal Computer running the UNIX System V Release 4.2 operating system with a 20 MB RAM and a 210MB harddisk with a transfer rate of 931.7 KB per second was used as the demonstration machine on which data was collected to validate the proposed predictive model. Statistical tests were applied to the outputs from the predictive model and it was found that there are no significant differences between the results of the prediction and the actual resource usage collected using the accounting tools of UNIX. Based on the validated predictive model, a capacity planning methodology is proposed which makes use of the predictions made by the model to plan for future changes and trends in the workload characteristics, as projected by the system administrator of the computer system. Use of the methodology may result in the performance tuning of the system when it encounters bottlenecks, or in the reconfiguration of the system when one or more system resource's capacity has been reached. The methodology is applicable both in the pre- and post-installation phases of the computer system's life cycle.