面向复杂城市场景的Sort-First并行绘制系统中负载平衡与性能分析

Load Balancing and Performance of Sort-First Rendering Clusters for Complex Urban Scenes

在高精度建模方法越来越丰富的背景下,三维城市场景往往具有较大的场景规模和较高的复杂性,因此,当单台计算机难以满足实时绘制的需求时,多采用更加适合超高分辨率大屏幕显示的Sort-First方式进行并行绘制。针对现有研究性能量化分析的缺乏,及自适应负载平衡方法的缺陷,本文通过Sort-First并行绘制系统的加速比模型,分析绘制节点数、显示分辨率、网络传输速度、任务划分等因素对系统性能的影响,以及将绘制时间和传输时间的和作为反馈,对现有负载平衡策略进行了改进。本文给出了系统达到最高加速比,即负载完全均衡理想情况下的一个量化的必要条件。最后,基于Sort-First的并行绘制实验验证上述分析。实验结果表明:本文的改进加强了各节点的负载平衡性,在一定程度上提高了性能。

With the rapid development of modeling methods, the scale and depth complexity of the 3D urban scenes have increased significantly. One of the methods to increase the system efficiency is to build multi-level 3D models. However, the simplification algorithms of 3D models are unsatisfactory in real-time rendering be- cause of the complexity of 3D urban models. Sort-First rendering clusters are broadly used when single PC can- not meet the requirements of real-time rendering, which is more suitable for high resolution and large screen ren- dering. The current researches are weak at quantitative analysis and self-adaptive load balancing. To overcome these difficulties, we studied the influencing factors of the performance by analyzing the speed-up ratio model of Sort-First rendering clusters, including node size, rendering resolution, network speed, task partitioning and so on. We also improved an existing load balancing method by using the rendering time and the transmission time as the feedbacks for adjusting the task partition instead of simply using the rendering time. Based on this, a nec- essary condition to achieve the highest speed-up ratio is proposed in an ideal situation that the load is completely balanced and the triangles are uniformly distributed on the screen. This quantitative expression of the condition may help us find out the bottleneck of the parallel rendering system and rationalize the configuration. We further point out that adding more rendering nodes might not help improve the performance if the rendering load is not very heavy, and the benefits of parallel system will decrease while the node size is large. A real-time parallel ren- dering experiment based on Sort-First was performed, which included 4 rendering nodes and used a large scale urban model. The results verify the above analysis and indicate that our improvement has increased the perfor- mance to a certain extent. Furthermore, our research could provide a reference for building Sort-First rendering clusters to reduce the abuse of the computer resources.