POEMS/MASC: Peformance-Oriented End-to-end Modeling System/ Model-based Adaptive Software Control

Directed by
M. K. Vernon,
UW-Madison Logo
V. S. Adve,
Univ of Illinois at Urbana Logo
R. Bagrodia,
J. C. Browne
Univ. of Texas at Austin Logo
E. N. Houstis,
Purdue Univ. Logo
J. R. Rice,
Purdue Univ. Logo
P. J. Teller,
Univ. of Texas at El Paso Logo
S. Wright
UW-Madison Logo

Current Topics

Project Sponsors

  • NSF Next Generation Software Program (September 2001 - August 2004)
  • NSF Next Generation Software Program (September 2000 - December 2001)
  • DARPA (July 1997 - December 2000)

Project Overview

This project brings together investigators with expertise in state-of-the-art applications, compilers, performance specification, performance modeling, and distributed system design/capacity planning to develop the tools needed for integrated design, development, evolution, and near-optimal adaptive run-time control of large high-performance applications running on distributed computational systems. Target applications include the cutting edge of computational science, such as complex stochastic optimization codes that are used to solve key organizational, economic, and financial planning decision problems that involve uncertainty. Target computational platforms include widely distributed heterogeneous "Grid" architectures running Grid middleware such as Condor, Globus, or Legion.

Grid platforms currently provide one of the most attractive environments for running large compute-intensive applications because the Grid resources are inexpensive, widely accessible, and powerful. Over time, the applications that are submitted using the Grid middleware are each given a share of the computational resources that are not used by higher priority computations. This enables an application to obtain large quantities of processing power relatively easily, although the the number and capabilities of the distributed hosts that will be allocated to the application is unpredictable and may vary during the course of the computation.

In contrast to previous work, the approach of the POEMS/MASC project is to develop model-based techniques and tools for near-optimal adaptive run-time control. That is, run-time control is based on high-fidelity performance models that control the execution of the application, so that the application adapts to its changing computational requirements as well as to the unpredictable changes in the allocated computational, communication, and storage resources, (nearly) optimally and/or to meet specified performance objectives.

For more information about how to get involved in the project, see the list of current topics.