Current POEMS-related Research
Collaborative Research: Performance-Driven Adaptive Software Design and Control
The fundamental concept underlying this research is that large-scale adaptive applications executing on complex dynamic distributed execution environments will require extensive runtime optimization if they are to take full advantage of both adaptive computational methods and "grid-like" execution environments. The goal of this research is to create a compiler and runtime supported model-based capability for runtime management and optimization of a dynamic application executing on a dynamic "grid-like" execution environment.
UTEP is focusing on developing models that can be used for application steering, in particular, for steering of adaptive applications.
Challenge
An adaptive application often has specified performance requirements. When such an application is executing on a distributed computational system meeting those requirements may be difficult due to fluctuations in the execution environment and mismatches between application behavior and/or available resources. Our challenge is to enable efficient design, development, evolution, and adaptive runtime control of complex adaptive applications executed on large distributed computational systems in order to meet specified performance requirements.
Goal
Our goal is to design and develop an advanced set of model-based tools for performance-directed control of an adaptive application executing on an adaptive grid.
Objective
Our main objective is to develop models that facilitate dynamic adaptation of application parameters based on changes in the execution environment and/or to meet a specific performance metric or constraint. A change in execution environment could consist of changes in local resources, e.g. processor loads, or in global resources, such as the addition or removal of one or more processing elements. A performance metric or constraint might be a real-time execution goal, or a specific accuracy for generated solutions.
Methodology
To reach our objective, we first conduct a parametric sensitivity analysis (a k-factor analysis) focused on the adaptable input parameters to the application. This analysis guides the future development of predictive performance models by quantifying the effect of individual parameters on a given constraint (e.g., execution time). The analysis also helps identify any parameter interdependencies that might be of importance when constructing a performance model.
Subsequent to the analysis of input parameters, mathematical modeling is done to generate a predictive model for the given constraint. This model, based on a desired constraint value, predicts the closest application parameter values that can achieve the needed performance to meet the constraint.