基于ROACH2-GPU的集群相关器研究——X-engine模块的设计与实现

A Research on the ROACH2-GPU-Cluster-based Correlator——The Design and Implementation of an X-engine Module

随着射电干涉技术的不断提升,干涉阵列规模越来越大,观测能力逐渐增强,但随之而来的是超大数据的实时处理问题。针对该问题,结合射电干涉仪相关器在数据运算和传输等方面的需求以及射电干涉阵列信号的特征,研制了一套基于图形处理器集群的通用相关器并用于"天籁计划"的数据处理:首先根据射电信号的关联计算特性,按频段将计算任务分配到不同图形处理器节点,并合理均衡各节点网络负载;然后由不同图形处理器节点独立完成各自的计算任务并将计算结果实时送往存储节点;最后按图形处理器集群通用相关器的设计方案成功安装部署系统并根据"天籁计划"一期的需求进行了性能测试。该图形处理器集群相关器计算性能约为理论峰值性能的46%;相对于传统方案的相关器,基于图形处理器集群的相关器具有开发周期短、可扩展性强、部署简单等优势。

As radio interference technology continues to improve,the scale of interferometric array becomes larger and larger. Its observation capacity also gradually increases. Yet real-time processing of big data becomes problematic. To tackle this kind of problem,this article takes the radio interferometer correlator＇s need of data computing and transmission,and the characteristics of the radio interferometric array signal into consideration and develops a set of generic correlator based on GPU cluster for the data processing work of ＂Tian Lai ＂project. First of all,considering radio signal＇ s characteristics of correlation calculation,computing tasks are assigned to different GPU nodes according to their frequency bands,and the network load on each node is properly balanced; then these tasks are completed by the corresponding nodes and the results are sent to the storage nodes in real time; finally,the whole system is deployed with reference to the data processing scheme of the GPU cluster correlator,and a performance test is carried out based on the first stage requirements of＂Tian Lai＂project. According to the results,the node computing performance of the cluster correlator has been speeded up： it is around 46% of the theoretical peak performance. Compared with the traditional correlator,the GPU-cluster-based correlator is superior owing to its short development cycle,strong scalability,simple deployment and other advantages.