A dynamic scheduling algorithm for divisible loads in grid environments

Divisible loads are those workloads that can be partitioned by a scheduler into any arbitrary chunks. The problem of scheduling divisible loads has been defined for a long time, however, a handful of solutions have been proposed. Furthermore, almost all proposed approaches attempt to perform scheduling in dedicated environments such as LANs, whereas scheduling in non-dedicated environments such as Grids remains an open problem. In Grids, the incessant variation of a worker's computing power is a chief difficulty of splitting and distributing workloads to Grid workers efficiently. In this paper, we first introduce a computation model that explains the impact of local (internal) tasks and Grid (external) tasks that arrive at a given worker. This model helps estimate the available computing power of a worker under the fluctuation of the number of local and Grid applications. Based on this model, we propose the CPU power prediction strategy. Additionally, we build a new dynamic scheduling algorithm by incorporating the prediction strategy into a static scheduling algorithm. Lastly we demonstrate that the proposed dynamic algorithm is superior to the existing dynamic and static algorithms by a comprehensive set of simulations.