This project has focused on two major issues: methods to analyze the application performance and methods to improve the application performance. The long term goal is to develop a methodology for systematically analyzing and improving application performance in the context of parallel scientific applications. To this end, this project has investigated techniques for analyzing and improving the parallel performance of three scientific applications: finite element analysis, transportation, and molecular dynamics. While these three applications have some common attributes (for example, the problem domains can be represented as graphs), they have significant differences such that this selection is representative of a large range of scientific applications. In addition to scientific applications, this work has focused on visual supercomputing environments. This past year, the focus has been on the finite element and transportation application.
In the area of finite element analysis, we have developed analytical models that allow the system characteristics to be considered for partitioning the problem mesh. Further, we have developed the initial version of a tool called PART, that considers the unique characteristics of distributed systems and the application to generate partitions. The finite element application has also been analyzed as an application running in a visual supercomputer environment. Within this context, we extended the work on lag models to identify the constraints on the execution of the supercomputer application.
Significant work has been with analyzing parallel transportation applications. Our analysis identified processor idle time as the major inefficiency in parallel shortest path computations, the computationally-intense step of traffic equilibrium problems. To address this inefficiency, we developed an adaptive method that resulted in a 63% reduction in parallel execution time for the traffic equilibrium applications as compared to methods that attempt to decrease the communication time. Further, we identified the network characteristics that must be considered to partition the network for efficient parallel execution.
The following significant results were obtained this past year in this project.
Accomplishment 1
Extended the work with lab models for virtual environments to include the simulation component of visual supercomputer environments. Conducted extensive experiments with various supercomputer applications that demonstrated the need to significant reduce the execution of fixed-size problems.
Accomplishment 2
Completed the development of the first version of a tool, called PART, that partitions a finite element mesh for distributed systems. Conducted some initial experiments that demonstrated that PART was able to achieve approximately 85% efficiency on a system composed of two IBM SP's, one located at Argonne and one located at Cornell.
Accomplishment 3
Conducted in depth-analysis of a parallel transportation application, in particular a traffic equilibrium application. The analysis aided in identifying the diserable network features that each partition should have for efficient execution on parallel machines. Further, we developed a method for reducing the idle time in parallel shortest path applications.