Solutions of Mesoscale Meteorological Mode

1. Demand analysis

Mesoscale meteorology is an important branch that develops rapidly in modern meteorological science. The atmospheric mesoscale motion it studies concerns detection and prevention of regional important disastrous weather. It reveals mesoscale observation facts on the one hand through mesoscale field test by applying a series of new detection tools such as satellite, radar, wind profiler and automatic observing station, on the other hand it carries deep simulation study and forecast test on process of mesoscale weather through mesoscale numerical mode. Along with the rapid development of computer technology in recent years, mesoscale numerical mode has become more mature, turning into an important research and application approach in mesoscale meteorology.

There are many mesoscale numerical weather forecast modes, among which MM5 is the mode that is in comparatively mature application at home and abroad at present, and is extensively used in domestic and overseas meteorological departments and relevant institutions. As a new-generation mesoscale mode developed on the basis of MM5 mode, WFR mode has now started to replace MM5 mode. GRAPES mode is a new-generation numerical forecast system developed independently by China Meteorological Administration; it is an advanced numerical forecast system commonly used in medium and small scale and large scale. Currently many domestic meteorological institutes have put GRAPES mode into research and application. AREMS mode, developed with the lead of Wuhan Rainstorm Institute of Chinese Academy of Meteorological Sciences, takes the processing method for complex terrains in domestic and overseas numerical modes into consideration, so it is relatively suitable for our country’s landscape feature and sets up specialized schemes for water vapor transfer and explicit way for cloud and rain in view of the importance and complexity of the process of water vapor; it forecasts preferably on rainstorm process in Jianghuai Valley. Other modes often used are LAPS and ARPS etc.

Features of mesoscale meteorological mode’s demand on computing environment are as follows:

（1）A huge computing amount

There is a spectacular calculated amount in mesoscale meteorological forecast mode (MM5, WRF, GRAPES and AREMS); meanwhile characteristics of weather forecast determine its high real-time performance, which requires the mode to finish calculation within designated time (usually not exceeding two hours). On the other hand, higher and higher requirements on accuracy of weather forecast are raised by people, while at present the accuracy of forecast is improved from hundreds of kilometers, a few tens of kilometers to several kilometers, which just substantially enhances the calculated amount of modes. Doubled precision of forecast will increase 16 times the calculated amount needed. The need of numerical weather report on calculating is impossible to be completed by a single CPU or an ordinary computer, so parallel computing should be used.There is a spectacular calculated amount in mesoscale meteorological forecast mode (MM5, WRF, GRAPES and AREMS); meanwhile characteristics of weather forecast determine its high real-time performance, which requires the mode to finish calculation within designated time (usually not exceeding two hours). On the other hand, higher and higher requirements on accuracy of weather forecast are raised by people, while at present the accuracy of forecast is improved from hundreds of kilometers, a few tens of kilometers to several kilometers, which just substantially enhances the calculated amount of modes. Doubled precision of forecast will increase 16 times the calculated amount needed. The need of numerical weather report on calculating is impossible to be completed by a single CPU or an ordinary computer, so parallel computing should be used.

（2）Extremely intense communication

Due to the fact that modes are concurrent software, and generally adopt the parallel computing of the finite-difference mode, amount of communication between each CPU is quite large when operating mesoscale meteorological forecast mode, as well as high requirement of the mode on communication’s performance. For example, communication of MM5 and WRF includes not only communication between domains of mother scope and nested scope, but also communication between different data partitions in each domain. Therefore, it requires high-performance communication network on high-performance computer.

（3）Strong real-time performance and timing running

Characteristics of weather forecast itself determine that it demands strong real-time performance. At the same time, forecasting system requires fixed and automatic operation, without manual intervention, commonly running in fixed two to four time intervals every day, with each time interval finishing running in two hours. These jobs are running at the same time every day, so it must be ensured that these modes can complete calculation on time.

（4） Main mode is where the calculated amount is

View from software’s processing flow, it’s usually divided into pre-processing, main mode and post-processing. Pre-processing includes data downloading and data assimilation etc., while post-processing mainly refers to generated products of graphical processing etc; pre/post processing usually requires not too much on computer. Main mode is the major part of the whole system, and also where the major calculated amount is, the part of which requires extremely high on computer’s performance.

2. Solutions

Specific to main characteristics of numerical modes in meteorological and oceanic fields: relatively large calculated amount of floating point, intensive communication, requirements of high stability and high I/O handling capacity on system, as to construct a mesoscale meteorological forecast mode system, Sugon thinks the following four points are very important when choosing basic hardware environment:

1. High performance, especially processing performance of floating point

2. High-performance network environment

3. High I/O bandwidth

4. High stability of the system

Below is the whole HPC cluster solution provided by Sugon Company about mesoscale numerical forecast:

A whole cluster hardware system includes subdivisions such as computing system, network system, storage system, administration login system and cabinet system; software system includes operating system, cluster management software, compiling development environment and meteorological operations software etc, as shown in the following table:

The solution has following advantages:

1. Computing blade supports the newest high-performance processor of Intel/AMD and leading high-density blade platform in the industry, providing incomparable high-performance computing power. By combining the newest computing network of 56Gb FDR Infiniband, it can fully meet the mode calculating needs of fine grid and large time span.

2. Cluster supports fault-tolerant function of calculation on hardware level and breakpoint calculation of numerical mode, and improves reliability of completion on time in numerical forecast’s key business.

3. Volume of massive parallel file system can be expanded to PB level and provide read-write bandwidth of high-performance IO, meeting demands of numerical mode’s requirements on storage system during processes such as data storage and processing, and mode calculation.

4. Sugon C1000 closed fluoride cold cabinet system provides an integration scheme with four subsystems of cabinet, air conditioning, power distribution and monitoring, achieving one-stop cabinet solution which is comprehensive, advanced and with functions meeting demands for users, so as to save space in computer room, improve refrigerating efficiency and reduce users’ capital investments.

5. Sugon meteorological operations software is based on the most advanced mesoscale forecast mode WRF, running automatically, and the whole process is without manual intervention from timing automatic downloading of background field data, pre-processing, mode calculation to post-processing, generating numerical forecast products.