Research area
Grids
Abstract
In the vision of grids as a computing utility, similar to the power grid for electricity and the telephone system for voice connections, users should be able to tap their resources and get guaranteed or at least predictable responses. However, in current grids, such guarantees and predictability are all but absent due to the lack of support for reservations in most local resource managers of the sites making up grids, the variability in the grid workloads, the variability in the available grid resources due to failures and to the background load of local users, and the variability of application runtimes. It is the purpose of this research proposal to answer the fundamental question to what extent grid schedulers can be built that do give performance guarantees to applications. We will therefore explore novel methods for performance predictions (data-mining techniques, confidence levels, failure inclusion), for application-level and Virtual-Organization-level scheduling (combining malleability and co-allocation in their interface to grid schedulers), and for grid schedulers (overdimensioning, job access control). We will assess these methods with statistical techniques, simulations, and experimentation in a real testbed.
Keywords
grid and multicluster scheduling, guaranteed response time
Partners
Leiden University (Lex Wolters)
Time frame
2006-2010
Funding
300 kEURO, NWO
Project web page
People involved
Dick Epema, Associate Professor