POEMS: End-to-End Performance Modeling of Large Heterogeneous Adaptive Parallel/Distributed Computer/Communication Systems

Collaborators: J. Browne (University of Texas at Austin, Project Management), V. Adve (Rice University), Rajive Bagrodia (UCLA), Olaf Lubeck (Los Alamos Nat. Lab.), Pat Teller (University of Texas at El Paso), Mary Vernon (University of Wisconsin)

Sponsor: DARPA

The POEMS project will create and demonstrate a capability for prediction of the end-to-end performance of parallel/distributed implementations of large scale adaptive applications. The POEMS modeling capability will span applications, operating systems including parallel I/O, and architecture. Research will focus on the areas where there is little convention wisdom available in the execution behaviors of adaptive algorithms on multi-level memory hierarchies and in parallel I/O operations. POEMS will provide:

• a language for composing models from component models,
• derivation of models of applications as data flow graphs from HPF programs
• a library of component models spanning from workloads to memory hierarchies and at levels of resolution ranging large grain data flow graphs to instruction streams and from probabilistic to fully deterministic.
• parallel execution of the models
• a knowledge base of performance data for commonly used algorithms parameterized for architectural characteristics.

The POEMS development is driven by modeling a full-scale Department of Energy ASCI (Accelerated Scientific Computing Initiative) application code executing on an ASCI architecture. This version of POEMS focuses on high performance computational applications and architectures. The intent is that the POEMS technology can be applied to other large complex dynamic computer/communication systems such as GloMo and Quorum.