基于Mark5B+GPU脉冲星观测系统

An Observation System for Pulsars Based on a GPU Architecture and a Mark5B

云南天文台40 m射电望远镜进行的脉冲星观测数据量巨大,必须实现数据的实时处理,否则将会产生海量的数据积压。为实现这一目标,采用图形处理器架构,对Mark5B数据进行解码、消色散、折叠等处理。实验结果表明,对以1 s 8 MB的实时采样,可以在0.51 s内处理完成,从而实现了实时处理的要求。首先介绍这一观测系统各部分的图形处理器实现,然后相对于传统中央处理器构架,对各部分的运算速度进行了详细的对比。针对时间开销最大的消色散部分,分析了单次傅里叶变换的数据量大小对执行效率的影响。从系统最终的输出轮廓和柱状图上可以看到实时处理的结果符合要求。最后对存在的问题和未来的工作进行了讨论。

Massive pulsar data are being generated through the 40m radio telescope of the Yunnan Observatories. A huge amount of data obtained by the telescope will stay unused if they are not processed in real time. To realize the needed real-time processing we have established a system of a GPU architecture to process data recorded by a MarkSB. The processing includes decoding, coherent removal of dispersion, and data-folding. Our test results show that data obtained in 1 second can be processed within 0. 51 second, which meets the requirements of real-time processing. In the paper we first introduce the implementation of each part of the observation system using the GPU architecture. We then compare the speed of each part of the system to that within a traditional CPU architecture. We particularly present the analysis of how the execution efficiency of a single Fourier-transform operation in the removal of dispersion is affected by the amount of data involved, as such operations are the most time consuming in the processing. Result profiles of tests with the system also show that our system meets the requirements of real-time processing. We finally discuss problems of the current system and future studies.