A real-time, GPU-based, non-imaging back-end for radio telescopes

Abstract

Since the discovery of Rapidly Rotating Transients (RRATs), interest in single pulse radio searches has increased dramatically. Due to the large data volumes generated by these searches, especially in planned surveys for future radio telescopes, such searches have to be conducted in real-time. This has led to the development of a multitude of search techniques and real-time pipeline prototypes. In this work we investigated the applicability of CPUs for such systems. We have designed and implemented a scalable, flexibile, GPU-based, transient search pipeline composed of several processing stages, including RFI mitigation, dedispersion, event detection and classification, as well as data quantisation and persistence. These stages are encapsulated as a standalone framework which can be used in offline mode, for processing archival data, as well as within an online application with additional real-time capabilities. The optimised GPU implementation of direct dedispersion achieves a speedup of more than an order of magnitude when compared to an optimised CPU implementation. We use a density-based clustering algorithm, coupled with a candidate selection mechanism to group detections caused by the same event together and automatically classify them as either RFI or of celestial origin. This setup was deployed at the Medicina BEST-II array in Italy, attached to an FPGA-based digital backend where several test observations were conducted. Finally, we calculate the number of CPUs required to process all the beams for the SKA1-mid non-imaging pipeline. We have also investigated the applicability of CPUs for beamforming, where our implementation achieves more than 503 of the peak theoretical performance. We also demonstrate that for large arrays, and in observations where the generated beams need to be processed outside of the GPU, the system will become PCIe bandwidth limited, with the attached CPUs spending most of the execution time waiting for I/O transfers. This can be alleviated by processing the synthesised beams on the GPU itself, and we demonstrate this by integrating the beamformer to the transient detection pipeline. We also analysed the beamforming computational requirements for SKA1-low and SKA1-mid, and demonstrate that CPUs are an inefficient architecture for this, with a very high running cost.