Grant: Discovery Grants Program - Individual

Recent important new challenges in system software for High Performance Computing are the omni-presence of multi-/many-core CPUs and the emergence of cloud computing. The proposed work focuses on solutions to these challenges in scheduling parallel HPC jobs on cluster/SMP servers, grids, and clouds. The purpose is to develop approaches with ideas which are feasible for practical application in Supercomputing Centers. The main goals of the work are predictability and quality-of-service (differentiated service) in job scheduling. Thus, the job scheduler will employ some form of preemption. Resource allocations can be flexibly defined as resource needs which may be met in either the time dimension (time sharing) or the space dimension (space partitioning) or a combination of both. Additionally, the resource shares may be adjusted in dependence on the job context. In this sense, the job scheduler is highly adaptive. The scheduling algorithms are accompanied by performance estimation. Response times strongly depend on the allocated resource shares. It is therefore important to know the performance resulting from certain resource-share allocations. This permits informed decisions. The performance estimation employs a combination of a model and machine learning of parameters. Therefore, the work will significantly support future server, grid, and cloud computing for parallel HPC jobs.