Sprungmarken

Servicenavigation

Hauptnavigation

Sie sind hier:

Hauptinhalt

SFB 531

 

Scheduling Strategies for Computational Grids

SFB The field of Computational Intelligence (CI) covers all sorts of techniques for subsymbolic (numerical) knowledge processing, such as the well known Fuzzy Logic, Neural Networks, and Evolutionary Algorithms as well as other approaches with lesser dissemination. Although CI techniques are widely in use, there still exists a large gap between theory and application. To close this gap the Collaborative Research Center (SFB) 531 has been founded at the University of Dortmund in 1997 as an interdisciplinary research institute. It is financially supported by the Deutsche Forschungsgemeinschaft (DFG). The scientific goals of the SFB are the investigation and improvement of the foundations, applications, as well as combinations of CI methods.

Overview

As the main topic of this reseach project we address the development of systematic methods and strategies for the creation of two-stage Grid-Schedulers for parallel machines.

The installation of Computational Grids enables users to access geographically distributed resources of different resource providers in an easy and transparent way. Similar to the power grid computational power becomes as easy and transparent accessable as electric power from a plug. Within a grid scenario users can submit their computational jobs, like for example simulations, to a grid management system. This management layer tries to discover suitable resources that fulfill the job's specific requirements. Furthermore, the infrastructure initiates the execution of jobs when the required resources become available.

The scheduling and resource management are particularly important within this context. An intelligent resource management provides an operant and thereforeeconomically beneficial application of the individual resources. An efficient operation of the system can only be achieved if the allocation of jobs to the different resources, like individual parallel machines or compute centers, is performed efficiently. However, so far no commonly accepted concept for a grid scheduling strategy exists.

Various strategies for parallel job scheduling have been developed and evaluated during the previous project C13 of the Collaborative Research Center 531. Within this project strategies and methods have been proposed that are able to created good schedules on the condition of multiple objectives defined by theresource owner.

For the future Computational Grid installations can be assumed which consists of several high performance computers located in compute centers or institutes. These compute centers will not be limited to participate in a grid environment. In fact they will join the grid permanently as an integral part of their usual operation. However, every resource provider will still control the local resources. Therefore, providers participate in the grid in order to improve their local job executions and increase their profit.

We extend the already existing scheduling strategies for parallel machines to a Computational Grid scenario. The main goal is the development of a Grid-Scheduler that supports the interaction of local resource management and the Computational Grid.

As the installation of Computational Grids is still a work in progress real workload traces are nowadays still unavailable. Therefore, we also address the development of a grid wokload model for our current project. This model is based on existing models for parallel machines which will be extended to a grid scenario. Furthermore, we will use this model for the simulation and evaluation of the grid scheduling strategies.

Research Activities

Project C13: Scheduling Strategies for Computational Grids of SFB 531 ''Design and Management of Complex Technical Methods of Computational Intelligence''
Information: Dipl.-Ing. Joachim Lepping