基于GPU的HOTPANTS并行优化
A Scheme of GPU-Based Parallel Optimization of the HOTPANTS Software Package
-
摘要: HOTPANTS是天文领域的重要工具,它主要是天文图像相减技术的具体实现,在凌日行星观测、引力透镜观测、寻找超新星以及发现天文现象的数据处理中发挥着重要作用。现阶段HOTPANTS经常被用在天文实时观测数据处理的流水线上,但是HOTPANTS在处理大规模的天文图像(7 K×7 K-10 K×10 K)时十分耗时。在分析HOTPANTS的性能瓶颈之后,提出并实现了基于图形处理器(Graphic Processing Unit, GPU)的HOTPANTS的并行优化方法。优化的HOTPANTS在处理大规模图像时,整体性能提升了2~3.5倍,卷积部分计算效率提升了6~13倍,并且通过图形处理器优化之后HOTPANTS的误差在天文领域是可以接受的。Abstract: The software package HOTPANTS is an important tool for astronomical image subtractions. It plays important roles in observation of transit planets, observation of gravitational lensing, supernova search, and discoveries of new astronomical phenomena. However, the original HOTPANTS package which is based on CPU processing is not suitable for processing massive real-time astronomical data. We analyze the problems limiting the performance of the HOTPANTS and modify it to be based on CUDA-GPU processing. The modified HOTPANTS package is 9 to 13 times faster than the original HOTPANTS package in performing image convolution. Efficiencies of complete image processing are improved by 2 to 3.5 times with the modified package. Our error analysis shows that results from the modified package are of acceptable accuracies like those from the original package.