EXPERIENCES ON DATA FLOW MODELS FOR DAG APPLICATIONS EXECUTED ON HETEROGGENEOUS AND CHANGING COMPUTATIONAL ENVIRONMENTS

摘要

We consider the problem of scheduling DAG applications onto heterogeneous and dynamic distributed computational platforms. These applications often require to compute and move a considerable amount of data among tasks. Most mapping methods focus on scheduling strategies which use the shape and static information of the DAG. They do not consider the mechanism through which communication of task results is actually achieved. We have found that ignoring this issue may negatively affect the performance of the application. In this paper, we explore two different models to allow the transfer of data among tasks, the PUSH model and the PULL model. Both models were implemented within the GTP (Global Task Positioning) model. GTP is a reactive scheduling method for DAG applications characterised by allowing rescheduling and migration of tasks in response to sig-ni cant variations in resource characteristics. We present the performance of the GTP model when using the PUSH or the PULL model.