A dynamic resource allocation and task scheduling strategy with uncertain task runtime on IaaS clouds

Abstract

Performing scientific workflows on IaaS cloud may face the problem of uncertain task runtime, meaning that tasks' real execution time are different from their given or evaluated time. So the pre-determined scheduling scheme can not perform as expected, causing that scientific workflows can not be completed within deadline or the total cost is far beyond users' budget. To address this problem, we proposed a new dynamic resource allocation and task scheduling(DRATS) strategy. At build-time, we use Path cut(PC) algorithm to generate a static task-Virtual Machine(VM) mapping scheme, and use task duplication(TD) algorithm to reduce the affect caused by their potential uncertain task runtime; At running time, when any task's executing time is inaccurate, we use least resource appending(LAR) algorithm to re-organize the mapping relationship between successor tasks and VMs, and add new virtual machine as cheap as possible. Experimental results demonstrate that, DRATS strategy can decrease the impact brought by uncertain task runtime, improve the probability of completing scientific workflows on time, and reduce the execution cost of scientific workflows effectively while satisfying the deadline constraint.