Modeling grids in (near) real time

Abstract

A Grid is a family of technologies for dynamically and opportunistically provisioning computing power from a pool of resources. Some experts believe that Grids are the next stage in the evolution of distributed systems. Enterprise-level Grids are beginning to be deployed, and researchers are investigating a number of novel Grid applications and services in Grid testbeds. Production-grade commercial Grids are expected to be large and widely-distributed systems. Grids are enormously complex computing systems. They have large numbers of geographically dispersed resources at their disposal, and they lack some characteristics of earlier computing systems that simplified the analysis of those systems - e.g., homogeneous and closely-coupled components, low-latency communications paths, global state maintenance, and deterministic controls. Performance analysts have very little understanding of the dynamic behavior of Grid components at different layers and at different time scalesy and of the complex interactions among Grid components -functionally, spatially, and temporally. We describe an architecture for a scalable, modular, configurable, plug-and-play, " Grid-in-a-lab " tool for modeling Grids, and for analyzing the performance of applications and middleware services offered on Grids. The tool addresses all of the components that are relevant to Grid performance - hardware, middleware, infrastructure, services, applications, traffic - And the complex interrelationships among the components. It uses a novel combination of modeling techniques - e.g., analytical models with closed form expressions, discrete event simulation, emulators for protocol stacks and resource scheduling, rare event simulation methods, and measurement-based optimization. The tool's components are modular and configurable so as to provide different degrees of fidelity over a range of time scales, and to provide insight about Grid dynamics and interactions among various components, applications, and services. © 2006 by Springer Science+Business Media, LLC.