Performance-based design of distributed real-time systems

This paper presents a design method for distributed systems with statistical, end-to-end real-time constraints, and with underlying stochastic resource requirements. A system is modeled as a set of chains, where each chain is a distributed pipeline of tasks, and a task can represent any activity requiring nonzero load from some CPU or network resource. Every chain has two end-to-end performance requirements: Its delay constraint denotes the maximum amount time a computation can tate to pow through the pipeline, from input to output. A chain's quality constraint mandates a minimum allowable success rate for outputs that meet their delay constraints. Our design method solves this problem by deriving (1) a fixed proportion of resource load to gave each task; and (2) a deterministic processing rate for every chain, in which the objective is to optimize the output success rate (as determined by an analytical approximation). We demonstrate our technique on an example system, and compare the estimated success rates with those derived via simulated on-line behavior.