基于多绘制管线的大规模并行体绘制性能优化技术

Performance Optimization Technique for Large-Scale Parallel Volume Rendering Based on Multiple Rendering Pipelines

针对数值模拟输出的大规模科学数据,体绘制方法为了刻画复杂物理特征,会进行高密度光线采样,但由此带来了极大的计算开销和数据增量。在国产自主CPU高性能计算机上,由于处理器单核的计算能力低于商业CPU,只能使用更多的处理器核来分担体绘制任务,从而引起了采样数据并行通信的可扩展性瓶颈。为充分利用国产自主CPU高性能计算机来高效完成体绘制任务,针对大规模并行体绘制提出一种基于多绘制管线的性能优化技术,通过多管线、多进程的两级并行模式来降低单条管线的并行规模。在大规模并行体绘制中,该技术将绘制目标图像划分成多个子区域,绘制进程则相应分组,每个进程组独立执行一条绘制管线,以完成图像相应子区域的绘制,最后再收集所有的图像子区域,形成完整图像并输出。实验结果表明,优化后的体绘制算法在国产自主CPU高性能计算机上可以扩展到万核规模,并能有效完成体绘制任务。

For large-scale scientific data output in numerical simulations,volume rendering methods inevitably perform high-density ray sampling to capture complex physical features,resulting in significant computational overhead and data increment.However,on domestic autonomous-CPU supercomputers,owing to the lower computing power of a single processor core compared to that of commercial CPU,more processor cores must be used to share volume rendering tasks;this leads to scalability bottlenecks in the parallel communication of sampling data.Full utilization of domestic autonomous-CPU supercomputers to efficiently complete volume rendering tasks is an urgent problem that needs to be solved.To address this problem,this paper proposes a performance optimization technique for large-scale parallel volume rendering based on multiple rendering pipelines;here,the parallel scale of a rendering pipeline is reduced by two-level parallelism:first,at the pipeline level,and then,at the process level.In large-scale parallel volume rendering after optimization,the rendered goal image is first divided into multiple sub-regions,and all rendering processes are grouped accordingly.Each process group then executes a rendering pipeline independently,and as a result,the corresponding sub-region of the image is produced.Finally,all sub-regions of the image are collected,and the whole image is output.Experiments demonstrate that the optimized volume rendering algorithm can scale to approximately 10000 processing cores on domestic autonomous-CPU supercomputers and can effectively complete volume rendering tasks.