Research And Implementation Of Energy-Efficient DAG Based Task Scheduling Based On Slack Allocation

The widespread use of high-performance computing(HPC)promotes the development of the cluster technology.The primary performance goal of cluster systems has focused on reducing the execution time of applications while increasing throughput.The performance goal has been mostly achieved by rational resource allocation and efficient task scheduling.However,with the development of high-performance computing,cluster systems provide powerful computing capabilities while also consume huge amounts of energy,which causes problems such as high operation cost,environmental impact,and low reliability.Therefore,high performance and low power consumption are two important needs in cluster systems.This paper mainly focuses on energy aware scheduling and actual implementation of DAG baed tasks in cluster systems,which contains two aspects:First,we develop an energy aware scheduling algorithm called EASLA for DAG based tasks in the context of Service Level Agreement(SLA)on DVFS-enabled cluster systems.In order to take full advantage of slack time to minimize energy consumption,the algorithm takes into account the effects of precedence constraints between tasks on slack allocation.It first finds the maximum set of independent tasks for each task to increase parallelism for using slacks,and then iteratively allocates each slack to the maximum independent set whose total energy reduction is the maximal.Radomly generated DAGs and real-world DAG applications are tested in our experiments.The experimental results show that our EASLA algorithm can save up to 22.68%and 12.01%energy consumption compared with the GreedyDVS and EvenlyDVS algorithms respectively in homogeneous environments,and 12.33%energy consumption compared with the EES algorithm in heterogeneous environments.Second,as the scheduling of DAG based tasks is mostly confined to theoretical study,we design and implement a prototype system for DAG task execution.The system uses two-level scheduling.A resource management and task scheduling system called SLURM acts as the upper scheduling,which provides the resources needed by task execution and performs the scheduling scheme provided by the sub-scheduling system.A sub-scheduling system acts as the underlying scheduling,which provides the scheduling scheme and submits it to SLURM.A claasical scheduling algorithm called HEFT and our proposed EASLA algorithm are integrated into the sub-scheduling system.This system puts theory into practice,which provides a way for the actual execution of DAG based tasks.It also implements energy-efficient scheduling of DAG based tasks in SLURM by combining the CPUfreq system.It makes up for the deficiencies of SLURM in the aspects of DAG based task scheduling and energy management.