Study Of An Implicit Parallel Operator Library For Ocean Modelling

Ocean models have become an important tool for oceanographers to study marine movements and environmental changes in the ocean.They have been widely used in the numerical simulation and forecasting systems of coastal estuaries all over the world.With the continuous development of high-performance computing and the increasing number of researchers engaged in marine research,the ocean models are moving toward higher resolution and the inclusion of more physical processes.In order to meet the increasingly sophisticated simulation needs and take full advantage of the hardware benefits of highperformance computing platforms,researchers have to develop higher-resolution ocean models on new computing platforms.The complex modelling work of these ocean models and the rapid development of parallel computing technology for supercomputers have created a huge gap,whereby researchers are unable to focus on scientific issues because of the complexity of the computing technology.In order to solve the challenges brought by large-scale parallel programming and performance tuning in ocean modelling,this thesis proposes an efficient implicit parallel operator computing framework for ocean modelling,which separates the ocean modelling from the optimization of parallel performance.The framework improves the programming efficiency of model developers in a parallel environment,allowing them to focus on the scientific issues,without being constrained by technical details such as message passing and parallel optimization.The main contributions of this thesis are as follows:(1)We abstract the original equations and discretization methods for ocean modelling.In addition,we propose field and operator to simplify the discretization representations.We also use an abstract staggered grid to encapsulate jumping rules and a grid increment to simplify the differential operations in the models.These basic abstracts ensure that the Open Array framework can decouple the work of ocean modelling from the work of parallel computing.(2)We design and implement an efficient computing library for ocean modelling.We also introduce computational graphs,lazy evaluation,kernel fusion,implicit parallelism,automatic code generation,just in time dynamic compilation and other technologies to ensure that the ocean model developed based on Open Array is easy to use,efficient and portable.(3)We develop a simple and efficient regional ocean model,where the lines of code are only 1/3 of the original regional ocean model.This achieves 91% parallel efficiency in strong scaling and 99% parallel efficiency in weak scaling with 4096 general CPU cores.It also exhibits excellent scalability on the heterogeneous multi-core supercomputer Sunway Taihu Light,achieving 85% parallel efficiency in strong scaling with 150,000 cores.