»   Workpackages

Efficient resource discovery is a key problem for Grid-type system processing. The nature of these systems because of their dynamics and difficulties in determining their future state causes significant problems in connection with system management and job scheduling. One of the difficulties that make the job scheduling quite complicated is the lack of sufficient information about users jobs. Especially there is too little knowledge about the job execution time on particular Grid-type system resources. It is obvious that acquiring the exact job execution time could significantly increase the efficiency of resource sharing algorithms used in standard resource management systems.

This workpackage is dedicated to creating a methodology of determining effective descriptors of job classes and methods of using the descriptors for dynamic determination of job execution time and job queuing time. Concurrently a method of predicting the future state of the resources state in a Grid system will be developed. The methods described above will be based mainly on deduction on the basis of previous (historical) job execution data and other system data like system logs. The research will concentrate on:

• User requirements analysis and definition of basic tasks in the system
• Preparing methods of knowledge discovery, acquisition and representation in an expert system module on the basis of the super-manager functionality analysis; designing and implementation of the module
• Expert system learning and verification of the knowledge base and inference rules
• Preparing, designing and implementation of a decision module on the basis of the super-manager functionality analysis
• Modules integration and examining system efficiency

The methods described above will be used mainly by the resource management system since it is expected that the efficiency and operating of the system will be significantly improved. Predicting job execution times and the future system state will allow to implement new versions of queue algorithms basing on backfilling and overbooking.