Resource Provisioning Methods For Cloud Workflow Applications

In cloud applications, operations with precedence constraints are used to describe complex business logics. The set of tasks with precedence constraints is called workflow. The work-load of cloud workflow applications usually changes as time changes. Users can rent (release) resources from (to) public clouds according to the real workloads. Precedence relations, da-ta locality, software locality and interval based pricing models make the resource provisioning for cloud workflows complex. It is desirable to design elastic resource provisioning algorithm-s according to the application characteristics to minimize the resource renting cost. The main contributions of this paper are as follows:(1) Workflow scheduling with non-shareable services. A complete-Critical-Path based Iterative heuristic (CPI), in which complete critical paths are iteratively constructed, is proposed. In the CPI, the critical path optimization problem is first transferred into the Multi-stage Decision Problem (MDP) with multiple objectives. Then, a dynamic programming based method is proposed for the MDP. Through optimizing the complete critical path iteratively, the resource renting cost of the whole workflow is minimized.(2) Workflow scheduling with shareable services. In traditional service computing, services are priced by per use. Rented services can not be reused by tasks of the same users, which leads to resource waste. Therefore, for services with interval based pricing models of public clouds, a Multiple-Rules based Heuristic (MRH), considering the interval based pricing models, data locality and software locality are developed.(3) indivisible batch-task based workflow scheduling. One batch task usually needs a large number of service instances (such as VM instances), In order to improve the resource uti-lization, a fast heuristic is proposed, which try to reuse the remaining fractions of rented time intervals. The workflow deadline is first divided into task deadlines based on the solution of a relaxed problem. Then batch-tasks are consolidated and scheduled to service instances according to two proposed time-slot selection rules.(4) divisible batch-task based workflow scheduling. In order to improve the utilization of re-sources further, tasks of each batch-task can be scheduled separately in some realistic appli-cations. A Unit-aware multiple Rules based heuristic (URH) is developed for the divisible batch-task based workflow. The workflow deadline is first divided into task deadlines by a minimized estimated-wasted-cost first method. Then tasks are schedule to services instances by a hybrid time-slot selection rule under the task deadline constraints.