Abstract

RTS is a new instrument designed to discover transient astronomical sources in the radio spectrum. To operate, it will require a high-performance radio transient pipeline, capable of processing hundreds of data streams in real-time, at a data rate of 36 GB/s. What makes the processing of these data streams challenging is the fact that radio signals received from space are dispersed by the interaction with the inter-stellar medium, and dispersion is a function of the distance between source and receiver. Because the distance of yet to discover objects is not known in advance, every data stream needs to be processed for thousands of trial distances, an extremely time consuming process. In this paper we introduce and describe a prototype for the ARTS radio transient pipeline. Our proposed pipeline is highly parallel and uses Graphics Processing Units as accelerators for its computational kernels. We test the pipeline on two different and widely used GPUs, the HD7970 from AMD and the K20X from NVIDIA, and show linear scalability and real-time performance on both. Using these performance results, we provide an estimate on the size of the system necessary to implement ARTS, and a lower bound on its power consumption. The results of this paper are also relevant in the context of designing the Square Kilometer Array.

Abstract

Finding new pulsars has always been a challenging problem, but this challenge is nowadays exacerbated by the increasing data rates of modern radio telescopes. Because of these increased data rates, traditional approaches to searching, based on storing data for off-line processing, are becoming unfeasible. Therefore, we propose a new pulsar searching pipeline that, by exploiting high-performance computing techniques, is able to process observational data in real-time. To achieve the real-time goal we parallelized all the steps of the pipeline to run on many-core accelerators, and used auto-tuning to adapt and optimize the pipeline for different platforms, telescopes, and searching parameters. In this paper, we test our pipeline on three different platforms: two Graphics Processing Units from AMD and NVIDIA, and an Intel Xeon Phi. Furthermore, we test it on three different scenarios, based on the operational parameters of three state-of-the-art telescopes. Results show that our pipeline can adapt to all tested platforms and scenarios, and achieves real-time performance and linear scalability. Because power consumption is a main concern for radio telescopes, and will be the main bottleneck for the construction of the Square Kilometer Array, we also measure the power consumed by our pipeline. By comparing the results obtained on many-core accelerators with the results obtained using a traditional multi-core CPU, we conclude that the accelerators can provide up to a factor 8 improvement in execution time, and up to a factor 6 reduction in power consumption.