Adaptive optimization of the performance of a virtual memory computer

It is known that the regulation of the degree of multiprogramming is perhaps one of the most important factors determining the overall performance of a virtual memory computer. In this paper we present an approach which differs some what from the approaches usually taken to regulate the degree of multiprogramming, which are mainly derived from the working-set principles.

We design a controller which will regulate the system in order to optimize a given performance measure. The controller is applied to a system where the critical resource is primary memory, and we are only concerned with systems where ineffective regulation leads to the phenomenon known as thrashing due to extensive paging activity.

In the first section, the dynamics of the system we wish to regulate are investigated using an analytical model. The system consists of a set of terminals and of a resource loop (CPU, secondary memory device, file disk) shared by the users. Using classical assumptions about program behavior (e.g., life-time function), the throughput of the RL is obtained as a function of the degree of multiprogramming n (number of users sharing the resources at a given instant of time) and of the system parameters. This result provides a greater insight of the “plant” we wish to control. The mathematical results are validated and extended with data from simulation experiments using a more detailed model (overheads and non-exponential assumption).

In the next section, a criterion called “dilatation” based on the utilization of the different resources is defined. From the analytical and simulation results of the first section, it can be shown that there exists a value n_o of the degree of multiprogramming which maximizes this criterion. The regulation of n to n_o is achieved by controlling the access of the users to the RL. The value of n_o is estimated in real-time through a continuous estimation of the two first moments of the criterion. Using these estimations, the decision of introducing or not a new user in the RL is taken whenever a user leaves a terminal or departs from the RL.

Extensive simulation experiments were conducted, where the implementation of the different functions of the controller have been thoroughly simulated. They have shown that the control scheme leaves to an improvement of the system performance in mean response time and resource utilization, and, overall, adapts in real-time the degree of multiprogramming to the characteristics of the users (the adaptation is performed in 4 sec. or so for a unit variation of the optimal degree of multiprogramming). A discussion of practical application of results ends the paper.