Adaptive Multi-level Workflow Scheduling with Uncertain Task Estimates
AUTHORS: Tomasz Dziok, Kamil Figiela, Maciej Malawski
Scheduling of scientific workflows in IaaS clouds with pay-per-use pricing model and multiple types of virtual machines is an important challenge. Most static scheduling algorithms assume that the estimates of task runtimes are known in advance, while in reality the actual runtime may vary. To address this problem, we propose an adaptive scheduling algorithm for deadline constrained workflows consisting of multiple levels. The algorithm produces a global approximate plan for the whole workflow in a first phase, and a local detailed schedule for the current level of the workflow. By applying this procedure iteratively after each level completes, the algorithm is able to adjust to the runtime variation. For each phase we propose optimization models that are solved using Mixed Integer Programming (MIP) method. The preliminary simulation results using data from Amazon infrastructure, and both synthetic and Montage workflows, show that the adaptive approach has advantages over a static one.