一种在异构系统中实现负载平衡的方法

提出了在异构系统实现负载平衡的区域分解算法和实现负载平衡的计算方法，利用它的负反馈性质解决了异构系统处理机计算速度测量误差造成的负载测量不准问题，并对处理机速度变化，速度测量误差、处理机数量、网格点计算量的分布等因素的影响进行了计算，结果表明本方法具有很强的平衡负载能力和较强的适应性；根据计算结果提出了解决模式网格点计算量不易测量问题的解决方案，并用扩散方程和模拟物理过程进行试验，试验表明这种方法是可行的，平衡负载的效果十分显著。     

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。     

中期数值预报系统T213L31在IBM/SP高性能计算机上的建立

在引进欧洲中期天气预报中心 (ECMWF) 的全球谱模式的基础上，通过对原模式的分析改造，首次以分布与共享相结合的方式在国家气象中心IBM/SP高性能计算机上实现了全球谱模式的高效运行。采用调整向量长度、优化程序设计、完善消息传递机制和实现MPI与OpenMP的混合并行编程等方法，减少模式的通信量、计算量和内存的使用量，提高了计算效率。实现了在T213L31分辨率条件下，10天预报可以在3 h之内完成，达到业务对时限的要求。建立了与T213L31全球谱模式相配套的最优插值（OI）并行处理分析系统，解决了由于观测站点在全球不均匀分布所带来的计算负载不均衡问题。在此基础上，实现了T213L31全球资料同化与预报系统并建立了相应的自动作业监控系统。     

区域三维变分同化中背景误差协方差的模拟

背景误差协方差(B)是变分同化中的一个重要部分,极大地影响同化系统输出的分析场.由于计算和指定B中有关统计量需要巨大的资料存储量和计算量,因此进行相关的研究较为困难.本文首先论述了B在变分同化中的重要性以及进行模拟的必要性;接着介绍了美国NMC方法的原理,并研究将其应用到区域三维变分同化中的方法;然后利用WRF模式生成的预报场差值集合对有关统计量进行了估计.揭示了以下结论:通过使用平衡变换和回归系数,控制变量被限制在较小范围内,保证了分析场的质量;流函数第一全局特征向量在200 hPa附近的最大分量,表示了急流层中强西风误差;流函数前五个全局特征向量在低层与中高层之间是负相关的;非平衡温度和相对湿度的特征长度尺度比流函数和非平衡速度势的值要小,说明它们是局地性较强的量.流函数和非平衡速度势的特征长度尺度随垂直模态数的增大快速减小,而相对湿度和非平衡温度的特征长度尺度随垂直模态数的变化较为平缓.     

中国对流层大气辐射平衡

本文根据辐射传输理论和某些物理近似,提出一个计算对流层大气辐射平衡的半经验半理论的气候学方法。使用十年平均气候资料,计算了我国对流层大气的辐射平衡及其有关分量,给出了这些量的分布特征,并简要讨论了影响我国对流层大气辐射平衡的主要因子。     

彭曼公式风函数的应用

引言当现有资料充足时,计算蒸散量(ET)的彭曼(Penman)综合公式是一个理论可靠且应用广泛的公式。这个公式把 ET 分为能量平衡和空气动力学两部分而把吸收太阳能和平流能的效应分开。空气动力学项包括两部分:第一部分是来自经验的风速函数,它与第二部分相乘。第二部分是饱和水汽压与环境或露点水汽压之差,这个差值叫饱和差。测定 ET 时,可将风函数与计算的饱和差进行验正。然而,计算饱和差的方法很多,其结果各不相同。因此,在实践中,计算饱和差的方法必须与所要使用的特定的风函数相一致。由于在饱和差的计算上概念不清,所以用这个公式时往往出现不确切和误差,计算饱和差的方法,至少有六种。因为风函数是用饱和差由经验得出的,所以风函数的大小和形     

区域大气环境容量测算及二氧化硫总量平衡分配研究

利用CALPUFF中尺度模型模拟计算出来的影响矩阵建立容量模型,使用专业的数学计算软件Lindo对容量模型进行计算,在综合考虑中部城市群区域的客观情况下,得到基于环境容量的辽宁中部城市群SO2排放总量平衡分配方案,为环境管理提供科技支撑。     

微型计算机及其发展趋势

所谓微型计算机,是指字长4～16位,运算速度比小型计算机慢一个数量级,但体积更小的一类计算机。微型计算机,有时也叫做微型处理机。1971年底,美国生产半导体器件的 Intel 公司首次制成的4004型微型计算机。这台计算机字长为4位,其中央处理装置采用并行处理方式,大小仅8mm×23mm,16个插脚,中央部分是一块大约3mm 见方的硅片,在这一小块硅片上,集积了大约2200个晶体管。后来,该公司又制成了8008型8位并行处理机及8080型机,Motorola 公司也制成了6800等机型。现在已经出现处理速度与小型计算机相当的16位并行微型处理机。     

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。     

一种用户均衡分配不确定性的计算方法

为解决目前交通分配中存在的不确定性问题,基于Wardrop用户平衡原理,利用起讫点（OD,Origin Destination）估计方法和Beckman交通分配模型,建立了一种交通分配不确定性计算方法.该方法分别以不同置信水平下的OD估计结果的上下限为输入量,然后利用Frank-Wolf算法求解交通分配模型,得到不同置信水平下的路段流量区间,以此量化交通分配问题中的不确定性.以南京市区域路网为研究对象进行案例分析,并采用宽度流量比R和无效覆盖率（Kickoff Percentage,KP）对模型结果进行评价,结果表明该方法可以得到路段流量的置信区间,量化交通分配的不确定性.     

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.     

自适应网格技术在数值模式中的应用研究 II.二维问题

用一个具有解析解的二维动力锋生成过程算例,对比自适应网格方案和固定网格方案的优劣,结果表明:同等误差要求下,固定网格方案的网格点数为自适应网格方案的3倍。从时间演化上看,自适应网格对温带气旋的预报,在同等误差要求下,可比固定网格的预报延长10小时以上。文章对加权系数选取及对解的影响进行了分析,从几个切面的计算结果展示了自适应网格对网格的优良安排并能抓住锋面特征结构。文章分析了光滑性、正交性对结果的影响。结果表明:网格的光滑性影响有一个优化值;在网格适当安排情况下,要适度考虑正交性。     

Using the spectral scaling exponent for validation of quantitative precipitation forecasts

This study evaluates the spectral scaling of a heavy rainfall event and assesses the performance of the Coupled Ocean/Atmosphere Mesoscale Prediction System (COAMPS) model in terms of the multiscale variability of rainfall in the temporal spectral domain. The event occurred over southern Malay Peninsula on 18 December 2006 and was simulated at high resolutions. 10, 5 and 1?min aggregate rainfall data from rain gauge stations in Singapore and simulated rainfall sampled at different evaluation points on 0.9, 0.3 and 0.1?km grids were utilized. The simulated and observed rain rates were compared via Fourier and wavelet analyses. A scaling regime was noted in the observed rainfall spectra in the timescales between 60?min and 2?min. The scaling exponent obtained from the observed spectra has a value of about 2, which may be indicative of the physics of turbulence and raindrop coalescence and might suggest the predominance of a characteristic raindrop size. At 0.9?km resolution, the model rainfall spectra showed similar scaling to the observed down to about 10?min, below which a fall-off in variance was noted as compared to observations. Higher spatial resolution of up to 0.1?km was crucial to improve the ability of the model to resolve the shorter timescale variability. We suggest that the evaluation of dynamical models in the spectral domain is a crucial step in the validation of quantitative precipitation forecasts and assessing the minimal grid resolution necessary to capture rainfall variability for certain short timescales may be important for hydrological predictions.     

上海区域降水集合预报系统的建立与运行结果的检验

以MM5模式为基础, 从预报模式的不确定性出发形成8个集合成员, 建立了上海区域降水集合预报系统。该系统实现从资料收集、资料处理、模式预报到预报结果处理与产品输出的全自动化, 于2005年8月1日开始业务运行, 运行稳定可靠。对系统8—10月的运行结果进行检验, 结果表明:集合预报系统对降水的总体预报效果尚可, 其中对量级小的降水的总体预报效果更好, 集合预报产品尤其是概率预报产品具有一定的参考价值, 但系统还存在发散度偏小的问题, 有待进一步改进。     

A comparative study of the performance of various vertical discretization schemes

Summary Most finite-difference numerical weather prediction models employ vertical discretizations that are staggered, and are low-order (usually second-order) approximations for the important terms such as the derivation of the geopotential from the hydrostatic equation, and the calculation of the vertically integrated divergence. In a sigma-coordinate model the latter is used for computing both the surface pressure change and the vertical velocity. All of the above-mentioned variables can diminish the accuracy of the forecast if they are not calculated accurately, and can have an impact on related quantities such as precipitation.In this study various discretization schemes in the vertical are compared both in theory and in practice. Four different vertical grids are tested: one unstaggered and three staggered (including the widely-used Lorenz grid). The comparison is carried out by assessing the accuracy of the grids using vertical numerics that range from second-order up to sixth-order.The theoretical part of the study examines how faithfully each vertical grid reproduces the vertical modes of the governing equations linearized with a basic state atmosphere. The performance of the grids is evaluated for 2nd, 4th and 6th-order numerical schemes based on Lagrange polynomials, and for a 6th-ordercompact scheme.Our interpretation of the results of the theoretical study is as follows. The most important result is that the order of accuracy employed in the numerics seems to be more significant than the choice of vertical grid. There are differences between the grids at second-order, but these differences effectively vanish as the order of accuracy increases. The sixth-order schemes all produce very accurate results with the grids performing equally well, and with the compact scheme significantly outperforming the Lagrange scheme. A second major result is that for the number of levels typically used in current operational forecast models, second-order schemes (which are used almost universally) all appear to be relatively poor, for other than the lowest modes.The theoretical claims were confirmed in practice using a large number (100) of forecasts with the Australian Bureau of Meteorology Research Centre's operational model. By comparing test model forecasts using the four grids and the different orders of numerics with very high resolution control model forecasts, the results of the theoretical study seem to be corroborated.With 8 Figures     

Modeling heavy precipitation in complex terrain

Summary A numerical prediction model is described which uses the full set of prognostic equations on a domain roughly the size of the United States with a 96 km horizontal grid resolution and six sigma-coordinate levels. Within this grid resides a nested domain of approximately 1000×1000 km with 24 km horizontal resolution. In this nested grid only modifications to the wind field by the better resolved terrain are considered on the lowest two sigma levels. The terrain effects necessitate adjustments in the location of these two sigma levels. Adjusted wind fields cause modifications in the mass and moisture divergence fields, hence in precipitation. These modifications are averaged into the appropriate meteorological fields on the larger grid.The algorithms used by our model allow continuous interaction between both grids with high computational efficiency.The relative importance of synoptic forcing and terrain is demonstrated for the cases of the Big Thompson, Colorado, flood of 1976 and the Cheyenne, Wyoming, flood of 1985.With 15 Figures     

A Cluster Space Calibration Scheme for Synthetic Wind Statistics Calculated with a Mesoscale Model and its Application to Regional Wind Climate Studies

Summary ?A scheme for calibrating an ensemble of wind fields computed by a mesoscale model in order to generate synthetic wind statistics is described. It is based on two main points. The first is to exploit the power of a mesoscale model to determine wind fields over complex terrain for different weather situations classified by cluster analysis. The second is to use all the information in the cluster analysis, i.e., the centroid values as well as the internal standard deviations of the clusters, to determine a cluster space distribution at each grid point in the model domain. The latter makes calibration possible if reliable measurements are available at the position of one of the grid points. The accuracy of the calibration is increased by splitting the cluster spaces into several parts. Combining both the modelled mesoscale wind fields and the method of split cluster spaces leads to a spatial transformation of the calibration from the calibration point to each grid point in the model domain. A validation of the scheme is carried out with measurements at grid points other than the calibration point and reveals remarkable improvements in the accuracy of the model wind statistics, especially with regard to wind speed distributions. Received October 8, 1998/Revised March 19, 1999     

High-resolution Simulation of an Extreme Heavy Rainfall Event in Shanghai Using the Weather Research and Forecasting Model: Sensitivity to Planetary Boundary Layer Parameterization

In this study,an extreme rainfall event that occurred on 25 May 2018 over Shanghai and its nearby area was simulated using the Weather Research and Forecasting model,with a focus on the effects of planetary boundary layer(PBL)physics using double nesting with large grid ratios(15:1 and 9:1).The sensitivity of the precipitation forecast was examined through three PBL schemes:the Yonsei University Scheme,the Mellor?Yamada?Nakanishi Niino Level 2.5(MYNN)scheme,and the Mellor?Yamada?Janjic scheme.The PBL effects on boundary layer structures,convective thermodynamic and large-scale forcings were investigated to explain the model differences in extreme rainfall distributions and hourly variations.The results indicated that in single coarser grids(15 km and 9 km),the extreme rainfall amount was largely underestimated with all three PBL schemes.In the inner 1-km grid,the underestimated intensity was improved;however,using the MYNN scheme for the 1-km grid domain with explicitly resolved convection and nested within the 9-km grid using the Kain?Fritsch cumulus scheme,significant advantages over the other PBL schemes are revealed in predicting the extreme rainfall distribution and the time of primary peak rainfall.MYNN,with the weakest vertical mixing,produced the shallowest and most humid inversion layer with the lowest lifting condensation level,but stronger wind fields and upward motions from the top of the boundary layer to upper levels.These factors all facilitate the development of deep convection and moisture transport for intense precipitation,and result in its most realistic prediction of the primary rainfall peak.