Comparison of the Execution Step Strategies on Scheduling Data-intensive Workflows on IaaS Cloud Platforms

Abstract

The IaaS platforms of the Cloud hold promise for executing parallel applications, particularly data-intensive scientific workflows. An important challenge for users of these platforms executing scientific workflows is to strike the right trade-off between the execution time of the scientific workflow and the cost of using the platform. In a previous article, we proposed an efficient approach that assists the user in finding this compromise. This approach requires an algorithm aimed at minimizing the execution time of the workflow once the platform configuration is set. In this article, we compare two different strategies for executing a workflow after its offline scheduling using an algorithm. The algorithm that we proposed in the previous study has outperform the HEFT algorithm.
 The first strategy allows some ready tasks to execute earlier than other higher-priority tasks that are ready later due to data transfer times. This strategy is justified by the fact that although our scheduling algorithm attempts to minimize data transfers between tasks running on different virtual machines, this algorithm does not include data transfer times in the planned execution dates for the various tasks of the workflow. The second strategy strictly adheres to the predetermined order among tasks scheduled on the same virtual machine.
 The results of our evaluations show that the best execution strategy depends on the characteristics of the workflow. For each evaluated workflow, our results demonstrate that our scheduling algorithm combined with the best execution strategy surpasses HEFT. The choice of the best strategy must be determined experimentally following realistic simulations, such as the ones we conduct here using the WRENCH framework, before conducting simulations to find the best compromise between cost and execution time of a workflow on an IaaS platform..