Energy-aware Task Scheduling Algorithms And Application For High-performance Computing

With the development of High-Performance Computing(HPC)technology,the performance of HPC has made a qualitative leap,but its energy consumption has also increased rapidly.Large-scale computing cluster systems consume more and more energy,causing various problems in terms of operating costs,environment,and system availability.At present,the power consumption of Super-Computing(SC)and HPC has reached megawatt level,and the top-1 SC “Summit” has reached 9.783 megawatt.Therefore,the energy consumption problem faced by HPC has become a critical challenge in the development of this field.This thesis focuses on the energy consumption and task scheduling of HPC systems,including energy-aware and energy-constrained task scheduling algorithms in homogeneous and heterogeneous computing environments.On this basis,for the large-scale computationally intensive applications,the proposed task scheduling algorithms are deployed in HPC systems and applied to the efficient simulation of the aircraft wing deformation.The main work and innovations of this thesis are as follows:(1)Energy consumption problem of HPC systems is focused and an Energy-Aware task Scheduling(EASLA)algorithm based on DVFS technology is proposed.The algorithm uses DVFS technology to assign different working frequencies to each task,thus effectively reducing the overall energy consumption of the computing system,while considering the Service Level Agreement(SLA)in terms of completion time and energy consumption.The EASLA algorithm first finds the largest set of independent tasks to increase the parallelism of slack scheduling.Then,the slack scheduling strategy is assigned to non-critical tasks to minimize energy consumption.Experimental results show that the EASLA algorithm can achieve a tradeoff between energy consumption and performance,which can effectively reduce the energy consumption of the computing cluster.(2)The energy consumption problem of large-scale heterogeneous HPC systems is studied,and an energy-aware task scheduling(HD-EASLA)algorithm is proposed for heterogeneous HPC systems.We respectively construct the heterogeneous computing node model,energy consumption model,and inter-node communication model.We design the task scheduling process under heterogeneous HPC systems in detail,as well as considering various constraints.The HD-EASLA algorithm assigns appropriate computing nodes to each computing task according to the energy consumption constraints of the application,and uses DVFS technology to adjust the working frequency of each computing node,thereby effectively reducing the overall energy consumption of the heterogeneous HPC system.The effectiveness of the proposed algorithm is evaluated by using task scheduling experiments on practical applications of the molecular dynamic coding program and the sonar data stream application.(3)An energy-constrained task scheduling(RSMECC)algorithm for heterogeneous HPC computing systems is proposed.The RSMECC algorithm is applied to applications in a HPC computing system,and scheduling tasks can be optimized under the condition of satisfying the given energy consumption constraints to minimize the scheduling length.A task pre-allocation mechanism based on energy consumption level is proposed,which provides a strict basis for the energy consumption constraints of task assignment journals.Extensive simulation experiments are performed to verify the effectiveness of the proposed RSMECC algorithm.Experimental results show that the proposed algorithm can predict the energy consumption of unallocated computing tasks and obtain an optimal task scheduling scheme under energy constraints.(4)The practical application of parallel task scheduling algorithms is studied.A HPCbased parallel mesh deformation algorithm is proposed and then applied in the application of aircraft aircraft wing deformation simulation.The parallelization algorithm of Radial Basis Function(RBF)and Recurrence Choleskey Decomposition Method(RCDM)is proposed respectively.According to the RBF and RCDM methods,the parallel mesh deformation algorithm is realized,and then is applied to the aircraft wing deformation and aerodynamic structural aeroelastic analysis.The experimental results show that the proposed parallel task scheduling algorithm can effectively improve the performance of mesh deformation.The work of this thesis achieves important research and practical application value.Especially in the era of high performance computing and energy saving and environmental protection,HPC and parallel computing resources are fully utilized,and various task scheduling algorithms are studied to effectively reduce the energy consumption of computing systems.This work also explore the application of HPC and task scheduling technology in the field of aviation,which lay a solid foundation for practical application research in other fields.