Results and Products

This project has the typical features of a significant experimental computer science activity. The work we propose is novel and as such has a major research component, which will be described in conventional journal and conference papers. However there are several major software system components that will produce artifacts of general value. These will consists of software realizations as well as innovative designs that will be documented by internal memos as well external publications.

All the software we develop will be in the public domain and highlights of our work will include:

• Development of two application emulators, which will emulate a total of 4-8 applications. The emulators will be simplified from real applications, but contain the essential communication, computation, synchronization and I/O characteristics that impact performance. The emulators will allow generic performance studies of important application classes.
• Development of a multi-level hierarchical application modeling framework, HLAM, where the products include design as well as a Java graphical user interface allowing interactive specification. HLAM will specify an application in terms of the smallest data blocks (called aggregates) needed to reliably understand performance.
• HLAM will also have an API to allow its generation from static and runtime compilation of application emulators.
• Development of a hierarchical machine specification capable of defining architectures expected in the next ten years, including those identified in the PetaFlop workshop studies.
• Development of a coarse grain simulator, PetaSIM, which will take HLAM as input, together with a machine specification, cost models and an execution script.
• Our activity will naturally produce reports describing and assessing the results of PetaSIM, which will validate both the approach and particular high level machine and application HLAM specifications.
• PetaSIM in its base form will be produced in both C and Java allowing standalone and applet execution. The system will be made available on a Web site with both the client download (Java) and higher performance (CGI or more modern servlet) server side implementation.
• Parallel implementation of PetaSIM is necessary for very large scale adaptive problems, and is a program option.