Distributed computing configuration: A combined user, software, and hardware model and analysis methodology.

摘要

This research uses the shared tools of Industrial Engineering and Computer Science to address hardware and software configuration selection for a distributed computer configuration. Using a single system model, we develop a unified method for selecting a machine to run each of the tasks in the application, for determining the maximum number of active copies of each task, and for approximating the mean response time for each of the users' activities. The objective is to minimize a performance metric that is positive when the target mean-response-time for any user activity is not met. We compare the results of this approach for calculating mean response times with simulation results. The traditional method of using simulation to evaluate potential configurations in a simple experimental design typically explores only a few possibilities with simulation run time ranging from minutes to hours. The method developed in this dissertation explores possible designs for a system with eight user activities using nine different tasks in less than five minutes.; This document: (1) Develops a consistent system model of a group of users, the distributed application they are using, and the computer network that runs the software. The model connects current research results in optimum assignment of tasks, modeling of computing resources, approximate solutions of queueing networks, and stochastic models of user behavior. (2) Extends the current techniques for optimum static assignment of tasks with deterministic service times to address models where service time is a random variable. This research also improves the speed of the existing algorithm for finding the best task and machine combination by ordering the search to first consider groups of tasks with the highest inter-task communications requirements. (3) Develops a heuristic for determining the optimum value for the maximum number of copies for each task given a system model and a task assignment. (4) Applies known principles in semi-Markov processes and queueing models requiring simultaneous resource possession to create an efficient iterative solution method for response time given a system model. (5) Develops new approximations for the response time calculation for both processor-sharing and first-come-first-served "multi-servers" for use in approximate mean value analysis. Unlike previously reported results, these approximations do not require additional storage between iterations and have not induced convergence problems in the approximate mean value analysis algorithm.