The Korean Integrated Model (KIM) System for Global Weather Forecasting

The Korea Institute of Atmospheric Prediction Systems (KIAPS) began a national project to develop a new global atmospheric model system in 2011. The ultimate goal of this 9-year project is to replace the current operational model at the Korea Meteorological Administration (KMA), which was adopted from the United Kingdom’s Meteorological Office’s unified model (UM) in 2010. The 12-km Korean Integrated Model (KIM) system, consisting of a spectral-element non-hydrostatic dynamical core on a cubed sphere grid and a state-of-the-art physics parameterization package, has been launched in a real-time forecast framework, with initial conditions obtained via the advanced hybrid four-dimensional ensemble variational data assimilation (4DEnVar) over its native grid. A development strategy for KIM and the evolution of its performance in medium-range forecasts toward a world-class global forecast system are described. Outstanding issues in KIM 3.1 as of February 2018 are discussed, along with a future plan for operational deployment in 2020.     

大规模数据并行问题的可扩展性分析

大规模数据并行处理的性能受到处理机数量、I/O速度、通信速度等多方面因素的制约。增加处理机数量或提高处理机的计算速度，可以提高计算机的整体处理速度，但是通信和I/O会成为影响并行效率的主要因素。为了综合分析这些因素对计算性能的影响，用一种比较典型的大规模数据并行的计算模型，具体分析了处理机数量、处理机速度与处理机间的通信延迟、通信速率以及输入输出速度之间的关系。得到了大规模并行机的通信和I/O性能与处理机速度与数量之间存在的关系。指出，增加处理机数量、提高单节点处理速度的同时，必须按照一定的关系相应增加节点间的通信性能和I/O性能。单纯以增加处理机数量、提高单处理机速度提高计算机峰值速度的方法会降低系统的计算效率，不能达到计算速度与计算机处理能力同步增长的目的。     

GRAPES全球模式MPI与OpenMP混合并行方案

随着多核计算技术的发展,基于多核处理器的集群系统逐渐成为主流架构。为适应这种既有分布式又有共享内存的硬件体系架构,使用MPI与OpenMP混合编程模型,可以实现节点间和节点内两级并行,利用消息传递与共享并行处理两种编程方式,MPI用于节点间通信,OpenMP用于节点内并行计算。该文采用MPI与OpenMP混合并行模型,使用区域分解并行和循环并行两种方法,对GRAPES全球模式进行MPI与OpenMP混合并行方案设计和优化。试验结果表明:MPI与OpenMP混合并行方法可以在MPI并行的基础上提高模式的并行度,在计算核数相同的情况下,4个线程内的MPI与OpenMP混合并行方案比单一MPI方案效果好,但在线程数量大于4时,并行效果显著下降。     

Development of an Observation Processing Package for Data Assimilation in KIAPS

A new observation processing system, the Korea Institute of Atmospheric Prediction Systems (KIAPS) Package for Observation Processing (KPOP), has been developed to provide optimal observation datasets to the data assimilation (DA) system for the Korean Integrated Model, KIM. This paper presents the KPOP’s conceptual design, how the principal modules have been developed, and some of their preliminary results. Currently, the KPOP is capable of processing almost all observation types used by the Korea Meteorological Administration (KMA) and some new observation types that have a positive impact in other operational centers. We have developed an adaptive bias correction (BC) method that only uses the background of the analysis time and selects the best observations through the consecutive iteration of BC and quality control (QC); it has been verified that this method will be the best suited for the KIAPS DA system until the development of variational BC (VarBC) has been completed. The requirement of considering the radiosonde balloon drift in the DA according to the increase of spatial resolution of the NWP model was accounted for using a balloon drift estimation method that considers the pressure difference and wind speed; thus the distance error was less than 1% in the sample test. Some kind of widely used methods were tested for height adjustment of the SURFACE observation, and a new method for temperature adjustment was outlined that used the correlation between temperature and relative humidity. In addition, three types of map projection were compared: the cubed-sphere (CS), equidistance (ED), and equirectangular (ER) projection for thinning. Data denial experiments were conducted to investigate how the KPOP affected the quality of the analysis fields in the three-dimensional variational data assimilation system (3D-Var). Qualified observations produced by the KPOP had a positive impact by reducing the analysis error.     

Computational Performance of the High-Resolution Atmospheric Model FAMIL

This paper describes the model speed and model In/Out (I/O) efficiency of the high-resolution atmospheric general circulation model FAMIL (Finite-volume Atmospheric Model of IAP/LASG) at the National Supercomputer Center in Tianjin, China, on its Tianhe-1A supercomputer platform. A series of three-model-day simulations were carried out with standard Aqua Planet Experiment (APE) designed within FAMIL to obtain the time stamp for the calculation of model speed, simulation cost, and model I/O efficiency. The results of the simulation demonstrate that FAMIL has remarkable scalability below 3456 and 6144 cores, and the lowest simulation costs are 1536 and 3456 cores for 12.5 km and 6.25 km resolutions, respectively. Furthermore, FAMIL has excellent I/O scalability and an efficiency of more than 80% on 6 I/Os and more than 99% on 1536 I/Os.     

Using urban effect corrected temperature data and a tree phenology model to project geographical shift of cherry flowering date in South Korea

Temperate zone deciduous tree phenology may be vulnerable to projected temperature change, and associated geographical impact is of concern to ecologists. Although many phenology models have been introduced to evaluate climate change impact, there has been little attempt to show the spatial variation across a geographical region due to contamination by the urban heat island (UHI) effect as well as the insufficient spatial resolution of temperature data. We present a practical method for assessing climate change impact on tree phenology at spatial scales sufficient to accommodate the UHI effect. A thermal time-based two-step phenological model was adapted to simulate and project flowering dates of Japanese cherry (Prunus serrulata var. spontanea) in South Korea under the changing climates. The model consists of two sequential periods: the rest period described by chilling requirements and the forcing period described by heating requirements. Daily maximum and minimum temperature are used to calculate daily chill units until a pre-determined chilling requirement for rest release is met. After the projected rest release date, daily heat units (growing degree days) are accumulated until a pre-determined heating requirement for flowering is achieved. Model parameters were derived from the observed bud-burst and flowering dates of cherry tree at the Seoul station of the Korea Meteorological Administration (KMA), along with daily temperature data for 1923–1948. The model was validated using the observed data at 18 locations across South Korea during 1955–2004 with a root mean square error of 5.1 days. This model was used to project flowering dates of Japanese cherry in South Korea from 1941 to 2100. Gridded data sets of daily maximum and minimum temperature with a 270 m grid spacing were prepared for the climatological normal years 1941–1970 and 1971–2000 based on observations at 56 KMA stations and a geospatial interpolation scheme for correcting urban heat island effect as well as elevation effect. We obtained a 25 km-resolution, 2011–2100 temperature projection data set covering peninsular Korea under the auspices of the Inter-governmental Panel on Climate Change—Special Report on Emission Scenarios A2 from the Meteorological Research Institute of KMA. The data set was converted to 270 m gridded data for the climatological years 2011–2040, 2041–2070 and 2071–2100. The phenology model was run by the gridded daily maximum and minimum temperature data sets, each representing climatological normal years for 1941–1970, 1971–2000, 2011–2040, 2041–2070, and 2071–2100. According to the model calculation, the spatially averaged flowering date for the 1971–2000 normal is earlier than that for 1941–1970 by 5.2 days. Compared with the current normal (1971–2000), flowering of Japanese cherry is expected to be earlier by 9, 21, and 29 days in the future normal years 2011–2040, 2041–2070, and 2071–2100, respectively. Southern coastal areas might experience springs with incomplete or even no flowering caused by insufficient chilling required for breaking bud dormancy.     

Beta Gyres in Global Analysis Fields

A three-component decomposition is applied to global analysis data to show the existence of a beta gyre, which causes Tropical Cyclone (TC) to drift from a large-scale environmental steering current. Analyses from the Global Data Assimilation and Prediction System (GDAPS) of the Korea Meteorological Adminis-tration (KMA), the Global Forecast System (GFS) of NCEP, and the Navy Operational Global Atmospheric Prediction System (NOGAPS) are used in this study. The structure of the beta gyre obtained in our anal...     

全球气象资料客观分析系统的分布式并行化

对已有的率行算法进行并行化，是一项很困难的工作。通过对全球气象资料客观分析系统率行算法的研究，提出了在MPP高性能计算机上的一种静态分配数据的分布式并行算法。该算法通过间隔选取分析盒子和模式格点纬圈行，将数据分配给不同的处理机实现分布式并行。该并行算法负载平衡好，并行效率高，而且并行化代价较低，具有良好的可扩展性。最后，给出了并行算法的性能测试结果。     

An Intercomparison of Rules for Testing the Significance of Coupled Modes of Singular Value Decomposition Analysis

This paper clarifies the essence of the significance test of singular value decomposition analysis （SVD）, and investigates four rules for testing the significance of coupled modes of SVD, including parallel analysis, nonparametric bootstrap, random-phase test, and a new rule named modified parallel analysis. A numerical experiment is conducted to quantitatively compare the performance of the four rules in judging whether a coupled mode of SVD is significant as parameters such as the sample size, the number of grid points, and the signal-to-noise ratio vary.
The results show that the four rules perform better with lower ratio of the number of grid points to sample size. Modified parallel analysis and nonparametric bootstrap perform best to abandon the spurious coupled modes, but the latter is better than the former to retain the significant coupled modes when the sample size is not much larger than the number of grid points. Parallel analysis and random-phase test are robust to abandon the spurious coupled modes only when either （1） the observations at the grid points are spatially uncorrelated, or （2） the coupled signal is very strong for parallel analysis and is not weak for random-phase test. The reasons affecting the accuracy of the test rules are discussed.     

一种适用于有限差分模式的负载平衡区域分解方法

分布式内存并行处理在数值天气预报等超大规模科学计算中已经得到了广泛的应用。中尺度模式由于分辨率高 ,计算量大 ,需使用更多的处理机进行并行运算。另一方面 ,由于复杂的物理过程的采用 ,增加了不同天气的计算量的不平衡。但是 ,目前所广泛使用的并行处理方法在处理机数量较多时不能很好地均衡计算负载 ,引起并行计算效率的降低。本文提出了一种新的非规则区域分解负载分配方法。并与已有的负载分配方法进行了分析试验对比 ,该方法能更有效地平衡负载 ,取得更好的加速效果     

Evaluation of Unified Model Microphysics in High-resolution NWP Simulations Using Polarimetric Radar Observations

The UK Met Office Unified Model(UM) is employed by many weather forecasting agencies around the globe. This model is designed to run across spatial and time scales and known to produce skillful predictions for large-scale weather systems. However, the model has only recently begun running operationally at horizontal grid spacings of ～1.5 km [e.g.,at the UK Met Office and the Korea Meteorological Administration(KMA)]. As its microphysics scheme was originally designed and tuned for large-scale precipitation systems, we investigate the performance of UM microphysics to determine potential inherent biases or weaknesses. Two rainfall cases from the KMA forecasting system are considered in this study: a Changma(quasi-stationary) front, and Typhoon Sanba(2012). The UM output is compared to polarimetric radar observations in terms of simulated polarimetric radar variables. Results show that the UM generally underpredicts median reflectivity in stratiform rain, producing high reflectivity cores and precipitation gaps between them. This is partially due to the diagnostic rain intercept parameter formulation used in the one-moment microphysics scheme. Model drop size is generally both underand overpredicted compared to observations. UM frozen hydrometeors favor generic ice(crystals and snow) rather than graupel, which is reasonable for Changma and typhoon cases. The model performed best with the typhoon case in terms of simulated precipitation coverage.     

GRAPES全球格点模式的并行计算负载平衡策略

随着高性能计算机技术的发展和应用,并行计算已成为保证数值天气预报模式业务运行时效的关键技术之一.目前高性能计算机计算能力已达到每秒千万亿次浮点计算.系统中处理器数量也早已达十万甚至更多,如此巨大的计算资源对应用软件系统的设计也提出了挑战.数值天气预报软件系统要充分利用高性能计算机提供的计算资源,必须依靠并行计算方法,这包括适合计算问题的可扩展并行算法的设计、合适的数据分配方案以及良好的任务负载平衡方案.作为中国新一代数值天气预报格点模式,GRAPES(Global and Regional Assimilation and PrEdiction System)设计的最终目标是一个科研/业务通用,区域/全球通用模式.作为一个格点模式,GRAPES的并行计算具有与欧洲中期数值顶报研究中心谱模式并行计算不同的特点,GRAPES的并行计算采用了经典的水平网格数据划分.但对于全球的GRAPES模式,由于采用拉格朗日差分方案,模式极地及附近区域格点与格点之间距离的减小.使得模式并行计算在采用简单的经纬网格划分方式实现时,必须考虑极地区域并行计算跨越多个处理器时导致的频繁通讯解决途径.本研究提出了利用消息传递组通讯实现全球格点模式并行计算的一种方法,其核心思想是将极点附近一定区域内的处理器按纬向划归不同的处理器组.文中还给出了该实现方法的任务分配算法,提出了改进的任务分配负载平衡方案.在中国气象局高性能计算机IBM-cluster1600上的测试表明,算法具有较好的可扩展性,其负载平衡方案改善了计算的绝对墙钟时间,使并行计算效率提高10%以上.模式的准业务运行结果表明计算墙钟时间基本可以满足数值预报业务的实时性要求.     

Cascade插值方法在GRAPES模式中的应用

基于半拉格朗日 (semi-Lagrangian) 方案的数值天气预报模式, 求解半拉格朗日轨迹上游点变量, 通常采用传统直线逐点拉格朗日多项式插值, 由已知模式格点 (欧拉网格点) 的数值插值获得。对于三维空间上游点的插值, N阶精度需要O(N3) 运算量。N增大, 运算量将大幅增加, 特别耗费计算机机时, 而采用Cascade插值法 (降阶插值法) 则只需要O(N) 运算量。它的显著特点是:用曲线代替直线, 通过一系列中间过渡网格点, 在曲线上用一维拉格朗日插值, 使得相邻拉格朗日格点或中间过渡点的插值不再是孤立的, 而且可以重复使用某些中间结果, 达到减少运算量的目的。将这种方法合理应用于GRAPES模式, 并根据模式的特点, 对Cascade插值过程中独立变量的距离分段计算, 从而有利于实现并行计算。计算结果表明Cascade插值法与传统直线逐点插值法相比, 计算效率平均提高约30%, 同时不降低精度。     

MARS软件在数值预报模式产品数据管理中的应用

该文简单介绍MARS (Meteorological Archival and Retrieval System) 软件的架构及组成部分,并基于MARS软件,介绍了数值预报产品数据管理系统设计与实现。通过在GRIB2数据段对应描述层次类型和气象要素的表格文件中定义新的层次类型和气象要素,在MARS软件解析GRIB2数据所需要查找的相关模板表格文件中定义气象要素对应的产品类型号、要素类别号、要素号、要素名称和要素名称缩写,说明GRIB2数据模板和代码表创建方法。以GRAPES模式产品GRIB2数据为例,实现应用MARS软件管理GRAPES模式产品GRIB2数据。结合存储管理TSM (Tivoli Storage Manager) 软件,研究开发MARS软件和TSM软件连接接口程序,实现数据的直接迁移和回调。以T639L60模式产品为例,通过测试检索性能,MARS软件检索数值预报产品速度超过当前国家级气象数据管理系统数据检索速度1倍,说明MARS软件检索数值预报产品数据具有时效优势。     

GPFS文件系统原理和模式IO优化方法

GPFS(通用并行文件系统)普遍应用于IBM大规模集群系统中,能够为并行应用程序提供高性能的I/O存取访问。该文介绍了GPFS的原理和工作机制,并以安装在国家气象信息中心的IBM高性能计算机系统为试验平台,初步测试了在实际环境中GPFS的应用情况,在应用程序中对文件访问进行性能优化的方式有2种:①由主任务负责全部文件访问,②通过MPI-IO在GPFS中实现并行文件访问。测试结果表明,GPFS可以提供很好的文件访问支持,但要想获得较高的性能,需要在应用程序中进行深入调试。     

Multi-resolution Hybrid Data Assimilation Core on a Cubed-sphere Grid (HybDA)

This study illustrates the characteristics of the data assimilation system at the Korea Institute of Atmospheric Prediction Systems (KIAPS), based on the cubed-sphere grid system. The most interesting feature is the use of spherical harmonic functions defined on cubed-sphere grid points, which makes it possible to control the allowable physical wavenumber for the analysis increments. The relevant computational costs and parallel scalability are represented. The multiple-resolution approach is a distinguishable aspect of this data assimilation system. The wavenumber, up to which the analysis is conducted, increases as the outer iteration progresses. This multiresolution strategy is based on an investigation into the change of spectral components of analysis increments. The multi-resolution outer-loop provides cost-effective analysis-improvement, by explicitly controlling the analysis increments entered into the observation operator. To utilize the high-resolution deterministic forecast as a background state, it is subtracted from the forecast ensemble, to produce ensemble forecast perturbation that is hybridized with static background error covariance. Based on the cycled analysis experiments, the higher-resolution deterministic forecast is shown to preserve the high-frequency feature of the analysis increment relative to the ensemble mean forecast.     

基于地理信息服务的海量气象信息网络发布

针对WebGIS在海量气象数据发布中的诸多不足,通过对地理信息服务的研究,结合气象数据本身所具有的特点,对栅格气象数据的发布采用了一种灵活高效的影像金字塔模型构建算法,结合服务器端动态生成切片的方式,有效减少了I/O操作和重采样计算,从而提高了系统的处理速度。在完成矢量气象数据发布的基础上实现了各区站气象信息的多级显示,实现了气象信息资料的共享和网络发布。实践证明,采用该方式能够实现在改变缩放级别时,地图数据的快速加载以及无缝拼接,为后续气象GIS平台的开发以及高效运行奠定了基础。     

Establishment of a multiple-Doppler radar wind retrieval system and its application to the analysis of mesoscale convective systems

A multiple-Doppler radar wind retrieval system is established by using a three dimensional variation method. The system consists of two parts. One is to interpolate reflectivities and radial velocities of nine Doppler radars of the Korea Meteorological Administration (KMA) into one analysis domain. The other is to retrieve three dimensional winds by minimizing a cost function that includes the following costs and constraints: the observed radial velocity cost, background wind cost, continuity constraint and smoothness constraint. In order to verify the performance of the system, retrieved winds are compared with observed winds obtained from five wind profilers of KMA. The performance of the system depends on the relative position to the baselines between Doppler radars. However, the performance of the system is enhanced when the number of overlaps among the radial velocities increases. The system is applied to the analysis of the evolution of a mesoscale convective system (MCS) on the Changma front on 1 July 2005. The analysis result shows that a new convective cell is developed by the convergence of the low troposphere winds at the organizing stage. The analysis of the vertical vorticity reveals that, among the two vorticity generation terms to be calculated utilizing the retrieved winds, tilting or twisting source dominates the divergence source in most convective regions. The strong downdrafts associated with the storm are produced on the downdraft branch of a meridional direction secondary circulation across the Changma front.     

PM10 data assimilation over south Korea to Asian dust forecasting model with the optimal interpolation method

A data assimilation (DA) system using ground PM₁₀ observation for Asian Dust Aerosol Model version 2 (ADAM2), which is the operational dust forecasting model of Korea Meteorological Administration (KMA), has been developed with the optimal interpolation (OI) method. The observations are provided by the PM₁₀ network operated by KMA. Three DA experiments are performed to simulate a dust event observed in Korea from 1 March to 31 May 2009 with different assimilation cycles of 24 (DA24), 12 (DA12), and 06 hours (DA06). 48-hour forecasts from the adjusted Initial Condition (IC) of dust concentration are compared with control simulation (CTL) and observation from independent stations. It is found that CTL simulates spatial patterns of dust emitted and transported associated with a developing low pressure system over the dust source regions quite well, compared with satellite measurement. However, it appears that there is considerable uncertainty in estimating the concentration of dust. With IC adjustment, the model simulates improved dust concentration, showing considerably reduced RMSE, particularly for the prediction within 12 hours of forecast. At the same time, it is shown that the time interval of DA affects the predictability of ADAM2, so that DA06 appears to have better predictability within a 12-hour simulation, reducing RMSE by 50% compared with CTL. This suggests that assimilating PM₁₀ to the dust prediction model using OI has the potential to predict air quality in Korea when the cycle of assimilation is sufficiently short.     

塔什干低涡天气过程的计算机识别

根据塔什干低涡的定义,利用NCEP／NCAR日平均再分析高度场、温度场资料,运用计算机编程自动识别、追踪塔什干低涡的中心位置、中心强度和持续时间,进而获取塔什干低涡天气过程的数据集。此计算机识别程序中有两个关键判断条件：一是某格点i的高度值小于其外围格点高度值,二是低涡范围内任意格点的纬向温度二阶导数大于零。实践证明该方法对塔什干低涡系统的识别准确且省时省力,同时避免了人工挑选的主观性误差。