Performance evaluation and prediction for large heterogeneous distributed systems

Heterogeneity is a trend setting phenomenon in today's world of networked computing. Combined with distributed computing, and large scale systems, such heterogeneous systems exhibit overwhelming complexity. This dissertation addresses performance issues, such as measurement, analysis, modeling, simulation, and configuration synthesis for such systems. Our study is experimentally validated on a novel testbed system, which consists of heterogeneous hardware units and operating system types, and provides transparent access to hundreds of users in an academic instructional environment. The main emphasis of this case study work is to arrive at a first order model, that renders successful results, by observing key behavior parameters of the systems under consideration. Such a model and methodology carry general applicability to other heterogeneous systems. A detailed study was conducted to define the user model of such a system, and arrive at fundamental building blocks for a variety of user communities and load patterns. We describe these results as a three-class user model structure. Numerous sets of performance measurements were designed and executed to characterize the heterogeneous distributed system under investigation. These measurements are required as input for our modeling process. They also carry a substantial value for future modeling purposes. A modeling effort was carried out, leading to the derivation of a queueing network model for such systems. This effort was particularly challenging due to the limited availability of low-level system parameters. Discrete event simulation was utilized to arrive at performance results, and we show good fit of modeling results to measured performance. Configuration synthesis algorithms were further developed and demonstrated for heterogeneous systems that we study. Given quantities are assumed to be the expected user workload in a three-class characterization, hardware costs, and delay constraints. We develop an optimization algorithm that uses performance design goals to arrive at the system configurations of choice. Finally, resource sharing issues are addressed by studying session assignment strategies, and the resulting global performance and optimization characteristics. This work also provides a basis for further research on load balancing ongoing at UCLA.