Over the last decade, a dramatic increase has been observed in the ability to generate and store data in the course of parameter studies, performance analysis, and software testing. We have developed the ZENTURIO experiment management system for performance and parameter studies, as well as for software testing on cluster and Grid architectures. ZENTURIO uses a novel directive-based language named ZEN to specify arbitrary application parameters and a broad range performance metrics. A large number of experiments are automatically generated based on the ZEN directives which annotate arbitrary application files. A graphical user portal enables the user to create, launch, and monitor the experiments. Moreover, a limited fault tolerance mechanism has been included to restart experiments in case of casual system, job scheduler, or application faults. Graphical tools are used to automatically display output and performance data across multiple experiments. In this paper we describe our experience with ZENTURIO for performance and parameter studies of a material science kernel, a 3-dim. particle-in-cell simulation, a fast fourier transform, and a financial modeling application. Experiments have been conducted on a SMP cluster architecture with different communication networks, using PBS (Portable Batch System) and GRAM (Globus Resource Allocation Manager) as job managers.

Top