【作者】 夏清；

【导师】 王光谦；

【作者基本信息】 清华大学 ， 水利工程， 2007， 硕士

【摘要】 现有的大尺度流域水文过程模拟大多基于分布式水文模型建立,然而要将分布式水文模型应用于较大尺度空间,势必要先将空间划分成下垫面条件相对均一的单元,且单元的数量随着精度要求的提高而成指数级增加,这对计算能力提出了很高的要求。另一方面,流域水沙过程模拟的很多应用,尤其像洪水预报,希望计算时间越短越好。要解决这个问题,一个办法是提高计算机的硬件水平,从而增加浮点运算速度,但是这一办法成本昂贵且对计算速度的提高空间有限。另一办法,基于现有硬件建立计算机集群采用并行计算,对当前资源的利用率高,且具有很好的扩展空间。本论文详细阐述了适应大尺度流域水文过程模拟特点和需求的集群计算技术及其实现。大尺度流域模拟的集群计算从逻辑上来看具有四个层次,最底层是模型定义层,定义了一些最基本的模型,如产汇流模型和产沙模型,可以按照需要增添新的模型;自顶向下第三层是计算层,这一层在河网编码的基础上按照任务并发性特点设计出高效的并行计算流程;第二层是监控层,对并行计算的过程加以监测,并依据系统负载对计算层实时控制;最高层是应用层,利用并行计算的特点结合GIS技术实现河网计算状态的直观显示以及“先算先得”的快速结果查询方法。具体实现时,在硬件上,集群采用主从模式架设,同时与数据库服务器通过高速以太网连接。在数据上,静态数据,如地形数据,使用Oracle数据库集中存储于数据库服务器;动态数据,如雨量数据,使用Access数据库分布存储于计算结点。软件上采用分块实现,多进程运行的方式。其中,模型定义层和计算层构成计算进程,它封装于系统内部,使用跨平台的C语言编写;监控层与应用层构成监控进程,它是与用户发生直接交互的部分,使用Visual C++编写。计算进程和监控进程之间通过一套信息传递机制进行通信。本论文将集群计算应用于黄河多沙粗沙区1977年9月1日至9月21日的产汇流模拟中,24个计算进程时加速比达到8,并行效率达到33%;计算性能在各方面表现优异;模拟结果与实测过程吻合较好。更多还原

【Abstract】 Most of current large-scale watershed simulation models are built on the concept of distributed hydrologic model, which divides the watershed into small units with similar attributes in underlying surface. In a given watershed area, the number of the units will increase in exponential speed when the resolution standard grows. This is a big challenge for computation capability. On the other side, in the practical applications of watershed simulation, such as flood forecast, the sooner the computation result comes out, the better for people to handle emergencies. To meet the requirement for a high computing capability brought up by these features, a simple way is to raise the computing capability of current hardware, however, it costs much while the extent to which the computing speed can be elevated is limited. Therefore, cluster computing method, which is effective and cheap, can be adopted into the computation in large-scale watershed simulation model. This thesis discusses how to develop a cluster computing system for the hydrological process simulation at large river basin scale.This cluster computing system has a hierarchical structure with four layers. The bottom layer is Modeling Layer, which aims to develop some basic hydrologic models, such as the runoff yield & confluence model and sediment yield model, while other models can be added into framework if needed. The third layer from top to bottom is Computing Layer, which aims to define a coding method for river network, assign and reassign tasks to maintain the Workload Balance, and use adaptive control to optimize the computing efficiency. The second layer is Control Layer, which aims to synchronize configurations and data in computing nodes, supervise the computing processes, and provide reports for analysis and optimization. The top layer is Application Layer, which shows an interface to a GIS platform. Through this interface, the tributaries that have been computed in a certain node are dynamically displayed in the current map. The interface also provides a way to find out real-time information, such as the runoff yield and sediment yield process, on a certain tributary. This cluster computing system is realized at hardware level, data level and software level. In hardware level, a cluster is built in master-slave pattern, and connects to database server through 100M Switch Ethernet. In data level, static data such as topography data are collectively stored in database server with stable Oracle Database, and dynamic data, such as rainfall data are distributively stored in local computer with flexible Access Database. In software level, the system is developed in two modules separately, and the modules are connected by multi-thread programming while communicated by a specific information standard.This thesis also uses the cluster computing system to simulate the runoff process of the sandy and grit area of Yellow River from September 1st to September 21st. The result shows that the speed-up reaches 8 and parallel efficiency reaches 33% when the computing nodes is 24, the computing performance is excellent in all ways, and simulated results are well inosculated the observed results.更多还原