OPTIMISING DATA-PARALLEL PERFORMANCE WITH A COST MODEL IN THE PRESENCE OF EXTERNAL LOAD

Computational environments, such as clusters and grids, provide a cost-effective platform for running computationally-intensive and data-intensive parallel applications. When such computational environments are shared, the demand for resources is irregular and so the load is unpredictable. We have been exploring a task mobile skeleton guided by a dynamic cost model, which encapsulates self-aware mobile control for the Master/Worker pattern of data-parallel computations and is able to move running tasks amongst the available processors. In this paper, we propose a dynamic scheduler guided by a performance cost model. This model enables a skeleton to anticipate future resource needs, be sensitive to the run-time loads and decide whether it would be better to serve the tasks elsewhere. Our experiments show that the skeleton is effective in dynamically relocating tasks in the presence of varying external loads to decrease the overall processing time on shared multi-core environment