Optimization techniques in computer system design and load control

Analytic modelling has proven to be cost-effective in the performance evaluation of computer systems. So far, queueing theory has been employed as the main tool. This thesis extends the scope of analytic modelling by using optimization techniques along with queuing theory in solving the decision-making problems of performance evaluation. Two different problems have been attempted in this thesis. First, a queueing network model is developed to find the optimal capacities and speeds of the memory levels in a memory hierarchy system operating in a multiprogrammed environment. Optimality is defined with respect to mean system response time under a fixed cost constraint. It is assumed that the number of levels in the hierarchy as well as the capacity of the lowest level are known. The effect of storage management strategy and program behaviour are characterised by the miss ratio function which, together with the device technology cost function, is assumed to be represented by power functions. It is shown that the solution obtained is globally optimal. Next, two adaptive schemes, SELF and MULTI-SELF, are developed to control the flow of jobs in a multiprogrammed computer system. They periodically determine the number of jobs from each class that should be activated to minimize the mean system residence time without saturating the system. The computation is based on the estimated system workload in the next interval. An exponential smoothing technique is used to reduce the error in estimating the values of the model parameters. The service provided to each class (specifically, the mean response time) may be adjusted by changing the weight associated with the job class. From our simulation results, the schemes appear to be both stable and robust. Performance improvement over some of the existing schemes (50%, L=S and the Knee criteria) is significant under some workloads. The overhead involved in its implementation is acceptable and the error due to some of the assumptions in the formulation and solution of the model are discussed.Science, Faculty ofComputer Science, Department ofGraduat