A Fully Lagrangian Dynamical Core for the Met Office NERC Cloud Model
A Fully Lagrangian Dynamical Core for the Met Office NERC Cloud Model
复制标题
英国气象局 NERC 云模型的完全拉格朗日动力学核心
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
D. Dritschel
中科院分区：
文献类型：
--
作者：
G. Gibb;S. Boeing;D. Dritschel
我们描述了欧洲气象局项目“A Fully Lagrangian Dynamic Core for the Met Office NERC Cloud Model”所开展的技术工作，该项目的目的是将“Met Office NERC Cloud Model”(Met O Mate Parcel In Cell)代码合并到Met O Met OneCe NERC云模式中，从而赋予其MPI并行性和超越单个节点的缩放能力。首先，我们描述了MODIfi阳离子给MONC，以使它是包裹感知的，然后我们概述了MPIC的物理在MONC中的实现。然后，我们提供了MPIC与新开发的代码的深入比较，并调查了它如何扩展到数千个MPI过程。最后，我们讨论了该规范的局限性，以及为缓解这些局限性而需要开展的未来工作。我们发现，对于较大的问题规模，代码可以扩展到数千个核心，尽管在规模上性能的主要限制因素是傅立叶变换例程。尽管代码的MPI性能很好，但OpenMP的性能很差，在12个线程上的加速比仅为3。总体而言，在单个节点上，新代码比MPIC执行得更好，每秒在每个核心上执行更多的包裹操作。至关重要的是，该项目中开展的工作使得可以模拟的问题规模比最初实施MPIC时要大得多。
We describe the technical work carried out in the eCSE project ‘A fully Lagrangian Dynamical Core for the Met Office NERC Cloud Model’, whose aim is to incorporate the ‘Moist Parcel in Cell‘ (MPIC) code into the Met Office NERC Cloud Model (MONC), thereby granting it MPI parallelism and the ability to scaly beyond a single node. First we describe the modifications to MONC in order for it to be parcel-aware, then we outline the implementation of MPIC’s physics into MONC. We then provide an in-depth comparison of MPIC with the newly developed code, and investigate how it scales to many thousands of MPI processes. Finally we discuss the limitations of the code, and future work to be carried out to mitigate these limitations. We found that the code can scale up to many thousands of cores for large problem sizes, although the main limiter of performance at scale are the Fourier Transform routines. Despite good MPI performance from the code, OpenMP performance is poor, achieving a speedup of only 3 on 12 threads. Overall, on a single node the new code performs better than MPIC does, carrying out more parcel operations per core per second. Crucially, the work carried out in this project allows considerably larger problem sizes to be simulated than was possible with the original implementation of MPIC.