Objective:
Our objective has been to develop a performance prediction technology that will be effective for modern high-end distributed systems. Key requirements of such a technology are the ability to model application performance using dynamic information and within a specific time-frame. Since adaptivity is a key paradigm for the grid, performance prediction models must be designed to
We are developing a new methodology -- Performance Prediction Engineering -- to address the problem of incorporating time-sensitive, dynamic and heterogeneous performance information into application performance predictions for metacomputing environments. This methodology involves the design and development of structural models (performance grammars) which allow for application execution performance to be represented by dynamically parameterized compositional models, the association of "Quality of Information" (QoIn) attributes (such as "accuracy", or "lifetime") with predictions and parameters, and the development of facilities for performance forecasting, i.e. facilities which provide dynamic performance information and predictions with associated quality of information attributes.
The confluence of these techniques results in a new definition of performance prediction that is dynamic and time-sensitive, and that carries a quantifiable measure of quality. Exposure of the complex but quantifiable relationship between application requirements and deliverable resource performance will lead to a an increase in the performance achievable by applications on distributed heterogeneous systems.
Recent Accomplishments:
In the last year, we obtained empirical verification that Quality of Information metrics can be used to enhance performance prediction in the context of dynamic scheduling. Enhancements to the Network Weather Service (a distributed performance forecasting system) developed during the project yielded more accurate resource performance predictions and were demonstrated to be fundamental in the achievement of application performance in grid environments through scheduling. Published experiments showed that "accuracy" QoIn values can be used to successfully capture the range of application execution behavior in dynamic environments and can be utilized to develop adaptive performance-efficient application schedules.
We developed improved quality of information techniques in the Network Weather Service (NWS) network monitoring facilities. (This software was deployed and demonstrated during SC98 in November by Rich Wolski and students).
Dynamically parameterized performance prediction models for prototype parameter sweep grid applications were developed by project member Dr. Henri Casanova. We are developing a prototype adaptive scheduler for this class of applications.
Current Plan:
This will be the last year of this contract. In the remaining time, we plan to accomplish:
Technology Transition
The results of this research is a strategy for developing dynamically parameterizable compositional models which can be used by application schedulers to improve the execution performance of grid applications. To develop this technology, we have developed software prototypes of the the schedulers which are being transferred back to the application developers and users, as well as enhancements to the Network Weather Service which are deployed at approximately 50 sites in the U.S. and foreign countries. These NWS sites include partner sites participating in NPACI, NCSA, and the I2DSI storage testbed, the ETL and Tokyo Institute of Technology in Japan, The University of Torino in Italy, the New University in the Netherlands, and the University of Lyon in France. Researchers at these institutions are using the NWS to study resource performance predictability in dynamic, shared computational environments. Of particular interest are dynamic resource allocation (I2DSI), on-line system fault diagnosis (NPACI), resource-to-desktop performance prediction (NCSA), and effective network simulation (ETL). These research efforts derive direct benefit from NWS research supported by the Performance Prediction Engineering project. NWS information and dynamic performance forecasts for selected participating sites can be obtained from http://nws.npaci.edu.