University of Tennessee
|Topic:||Computational Grids: Aggregating Globally Distributed Resources for Performance|
|Date:||Friday, March 30, 2001|
|Place:||Gould-Simpson, Room 701|
The continuing proliferation of network connectivity brings with it the possibility of building applications that can use globally distributed resources to achieve high performance levels. The Computational Grid is an emerging paradigm for supporting such applications in which programs draw computational "power" from a global resource pool the same way appliances draw electrical power from a power utility -- seamlessly, ubiquitously, and anonymously.
In this talk we will present a novel approach to realizing the Computational Grid vision. To enable Grid programs to adapt to fluctuating resource performance and availability, we have developed the Network Weather Service (NWS). The NWS is a scalable and robust distributed system for gathering periodic performance measurements from a widely dispersed pool of resources. It also uses a suite of fast and adaptive time-series models to forecast future levels in near-real time. Using NWS forecasts to enable adaptive scheduling, we have developed EveryWare -- a software toolkit for building globally distributable high-performance applications. We present our experiences using EveryWare to deploy a computational mathematics application across four continents as part of a demonstration at a national supercomputing conference.
Finally, if Grid programs are to adaptively acquire and release resources according to load and availability, it is possible that resource allocations will oscillate as independent competing schedulers constantly adjust to load changes they induce. To analyze the stability and efficiency of the Grid as a whole, we present a framework for building Computational Grid resource economies called G-Commerce. Using G-commerce simulations, we show that a computational commodities market out performs auction-based methods as a global resource allocation strategy in terms of equilibrium and price stability.
Taken together, these research efforts compose a new and comprehensive approach to achieving the robust, adaptive, and ubiquitous computing goals of the Computational Grid.