基于层次化调度策略的渲染作业管理系统的研究与实现

发布时间：2018-03-05 14:26

本文选题：动态负载均衡　切入点：作业调度　出处：《山东大学》2012年硕士论文　论文类型：学位论文

【摘要】：渲染(render)是将三维几何模型经过一系列的计算,主要是附加一定的材质、纹理及色彩,并加上光源,得到具有真实感效果的二维图像的技术,在许多领域,如建筑设计的内部预览、汽车制造、广告设计,特别是在3D动漫电影产业中具有重要作用。在较大规模的三维动画制作过程中,由于渲染计算量非常大导致使用普通PC机进行渲染所需时间往往让人难以忍受。其次,由于在整个渲染过程中存在着海量的数据存储与访问,使得渲染的I/O瓶颈也成为影响渲染服务质量的关键问题。因此,如果没有机群渲染技术的协助,那么三维动画制作是难以完成的。机群渲染系统也称作渲染农场,它是指由许多运行渲染引擎的计算机组成的,在机群渲染管理软件的统一调度与分配下,利用机群的优势,协调工作进行同步并行计算,以此来完成所分配的渲染作业。由于渲染计算任务之间的联系非常少,因此并行计算的加速比几乎可以达到线性。许多渲染应用服务提供商,如renderbus、炫我、Rebus Faem、RenderCore等都使用机群渲染技术目前国内在动漫渲染领域所使用的机群渲染管理软件种类繁多,其中不乏有开源软件以及商业软件,但是商业软件价格昂贵且并行效率不高而开源软件又相对不够成熟。机群渲染涉及多种技术,系统的组成、调度以及管理都十分复杂,因此,开发一个功能全面的、并行效率高的针对于高效能计算平台的渲染管理软件有很高的应用和研究价值。本文的主要工作是在现有的作业管理系统(开源软件torque)之上又设计并实现了一个针对渲染应用的作业管理系统,它包括有用户管理,数据管理以及作业管理和资源监控四个基本模块并且以portal的方式向用户提供了一个简单友好的使用界面。同时为解决在机群环境下渲染服务面临的机群环境动态性,效率发挥以及资源管理器的重用等基本问题,本文提出了一个面向渲染的层次化的调度策略,它将渲染调度划分成了作业选择,资源选择以及负载均衡三个层次,层与层之间相互独立。这样使得我们能够较好地整合以及重用现有的机群管理软件,同时通过负载均衡,避免了性能最差的节点延长作业执行时间的可能性。论文的工作较系统地实现了在机群环境下对渲染作业进行调度和管理的基本需求,支持场景上传、作业提交、作业管理与监控、执行结果下载、渲染队列监控,实现了一个基本的面向高性能计算平台的渲染作业管理系统,具有良好的实用性和可扩展性。
[Abstract]:Rendering renderis a technology that combines 3D geometric models with a series of calculations, mainly by adding certain materials, textures and colors, and adding a light source to obtain two-dimensional images with realistic effects in many fields. Internal preview of architectural design, car manufacturing, advertising design, especially in the 3D animation film industry plays an important role. In the process of making 3D animation on a large scale, the time required to render with ordinary PC is often unbearable because of the huge amount of rendering computation. Second, Due to the massive data storage and access in the whole rendering process, the I / O bottleneck of rendering becomes a key problem affecting the quality of rendering service. Therefore, without the help of cluster rendering technology, So 3D animation production is difficult to complete. Cluster rendering system, also known as rendering farm, is composed of many computers running rendering engines. Under the unified scheduling and distribution of cluster rendering management software, the advantage of cluster is utilized. Coordinate the work of synchronous parallel computing to complete the assigned rendering job. Because of the very few links between rendering computing tasks, the speedup of parallel computing can be almost linear. Many rendering application service providers, For example, renderbus, Rebus Faemer, RenderCore, etc., all use cluster rendering technology. At present, there are many kinds of cluster rendering and management software used in the field of animation rendering in China, including open source software and commercial software. But the commercial software is expensive, the parallel efficiency is not high and the open source software is relatively immature. Cluster rendering involves a variety of technologies, the composition of the system, scheduling and management are very complex, so, the development of a comprehensive function, The rendering management software with high parallel efficiency for high performance computing platform has high application and research value. The main work of this paper is to design and implement a job management system for rendering application on the basis of the existing open source software torque. it includes user management. The four basic modules of data management, job management and resource monitoring provide users with a simple and friendly interface in the form of portal. In this paper, a hierarchical scheduling strategy for rendering is proposed, which divides rendering scheduling into three levels: job selection, resource selection and load balancing. This enables us to integrate and reuse the existing cluster management software and to avoid the possibility of the worst nodes extending the job execution time through load balancing. The work of this paper realizes the basic requirements of scheduling and management of rendering jobs in cluster environment, supports scene upload, job submission, job management and monitoring, results download, rendering queue monitoring, etc. A basic rendering job management system for high performance computing platform is implemented, which has good practicability and expansibility.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP311.52

【参考文献】