problem
Grid computing model is batch-oriented and does not support a number of import
use-cases.
Due to scale and complexity of the
infrastructure the reliability and performance of the Grid is varying in time. The
Quality of Service is variable and insufficient.
use cases
- sequences of jobs: parameter sweep and iterative applications (xmipp)
- unbalanced/unknown distribution of job duration (itu)
- man-in-the-loop: interactive batch applications (HEP analysis)
- sharing resources: local + grid
infrastructure
- LCG monitoring snapshot
- statistics
Quality of Service
- monitoring statistics, variations,...
- predictibility, shortened turn-around time...
user-level scheduling
- automatically deployed
- useful for non communication-intensive parallel applications
- dynamic load balancing, error recovery, ...
success stories
*
*
research roadmap
- define the Quality Of Service parameters such as predictability
- define the time-window model for grid computing (also taking into account local resources)
- investigate the impact of user-level scheduling on the fair-share
-
--
JakubMoscicki - 05 Oct 2006