Parallelization And Optimization Methods For Atmospheric Modeling On Heterogeneous Systems

Climate change is shadowing negative effects on both natural and social systems by causing destructive weather,damaging human health and reducing agricultural production.As one of the most important component of climate modeling research,atmospheric simulation has attracted much attention in high-performance computing.Developing more accurate and efficient atmospheric models to better predict extreme climate events has always been an important task and hot research domain of scientific computing.As atmospheric simulation is complex,extremely computing power is needed.Ever since the first numerical atmospheric program ran on ENIAC(the first electronic computer all over the world),atmospheric models have been among the major users of computing resources.With the increasing demand for simulation programs in the dimensions of model accuracy,forecasting scale,and data assimilation,the requirements for computing power of the atmosphere models is growing in a fast pace,and the development of climate model needs to be combined with the most advanced supercomputer of the era.On the other hand,constrained by the physical limits such as heat dissipation and power consumption,going purely for clock speed is no longer the best strategy in processor design.To meet the growing demand of computing power,many-core accelerators are developed and adopted on the leading-edge supercomputers in recent years and contribute more than90% computing power to these supercomputers.In the world's fastest supercomputer list(known as the Top 500 list)released in November 2018,heterogeneous architecture is adopted on all of the top seven supercomputers.After decades of development,the optimization and deployment of atmospheric models on homogeneous supercomputers(supercomputers with only one kind of processor)has been well studied.As homogeneous supercomputers are not able to meet the demand for performance growth,some models have gradually deployed to heterogeneous systems.In this work,taking the mainstream processors(POWER CPU,Nvidia GPU,Sunway processor)that adopted in the world's top three fastest supercomputers as example,we are aiming at proposing the optimizaiton methods and skills on heterogeneous systems,and analyze the similarities and differences between the optimization strategies on different processors in three levels: basic operator,kernel algorithm and complete model.The major contributions include:· proposing a unified heterogeneous domain decomposition method and computation-communication overlap mechanism based on all kinds of heterogeneous systems(namely on-chip and off-chip heterogeneous systems),thus to comprehensively demonstrate the process-level parallel methods and realize the efficient scheduling strategies of computing tasks;· taking IBM POWER CPU as example to demonstrate the optimization techniques for the key kernels of atmospheric model;proposing a performance tuning framework to facilitate the optimization based on CPU side,following by a set of hardware-software co-design(which fully combines the algorithm feature and hardware character)to fully unleash the performance potential of hardware;· based on the features of four generations of GPU platforms(Fermi,Kepler,Pascal and Volta),proposing a set of hardware-oriented performance tuning techniques for atmospheric programs(SWE and Euler solver)and get 3-6 times speedup at the first stage;following the novel tuning techniques(customized memory buffer design and inner-thread rescheduling)based on hardware-software co-design are adopted and an averaged 50% further speedup is achieved on different platforms;· aiming at optimizing commonly-used operators for atmospheric simulation based on SW26010 processor,besides properly adopting the hardware-oriented tuning techniques,we further proposing an optimization library that combined with two kinds of customized optimization strategies to fully unleash its performance potential and facilitate the performance tuning process;combining the deploying and optimization experience on Sunway and other platforms,the next generation of high-parallel atmospheric model towards exscale is developed in cooperation with atmospheric scientists.