SCOJO - Share Based Job Coscheduling With Integrated Dynamic Resource Directory in Support or Grid Scheduling

摘要

Grid environments currently use mere batch scheduling or assume jobs to be started immediately. We apply a combination of batch scheduling and time sharing and propose an approach -which assigns certain start times and certain shares based on integrated resource management. Employing loosely coordinated coscheduling, provides good chances for latency hiding and high flexibility in executing jobs across multiple sites. Our approach takes potential slowdowns or speedups from coscheduling into consideration. We support such scheduling by a dynamic directory service. Current directory services for grid computing provide static machine descriptions plus some dynamic information like CPU load as obtained via NWS. However, in other work, we have shown that the detailed application characteristics matter in coscheduling. Furthermore, predictability is currently limited to short-term extrapolations from the recent history. We develop scheduling plans and consider site workload and application characteristics, while not disclosing such information to other users or sites. This approach enables higher-quality decisions and longer-term predictability. We demonstrate our approach on the basis of simulations and simplified real coscheduling tests, run on an SMP server.