On the Scalability of Demand-Driven Parallel Systems

. Demand-driven systems follow the model where customers enter the system, request some service, and then depart. Examples are databases, transaction processing systems and operating systems, which form the system software layer between the applications and the hardware. Achieving scalability at the system software layer is critical for the scalability of the system as a whole, and yet this layer has largely been ignored. In this paper, we characterize the scalability of the system software layer of demand-driven parallel systems based on fundamental metrics of quantitative system performance analysis. We develop a set of sufficient conditions so that if a system satisfies these conditions, then the system is scalable. We further argue that in practice these conditions are also necessary. In the remainder of the paper, we use the necessary and sufficient conditions to develop a set of practical design guidelines, to study the effect of application workloads, and to examine the scalabil..