GPU-based acceleration searching

Abstract

Since the discovery of globular cluster pulsar, a number of exhaustive searches were performed, yielding over 140 globular cluster pulsars residing in at least 26 different clusters, 74 of which are found in binary systems. Searching for a priori unknown pulsars entails that in conducting any search procedure a wide range of parameters including the pulsar periodicity and the distance to the pulsar has to be taken into consideration. Searches for pulsars in binary systems are even more challenging due to the orbital motion smearing caused by Doppler-shift changes of binary motion. Consequently, binary pulsars show periodic changes in their pulse frequency derivative, causing a reduction in the Signal-to-Noise (S/N) ratio. Correlating the spectral components with an inverse frequency and a complex conjugate kernel template response enabled the recovery of the coherent response of the signal from the orbital motion caused by the acceleration of binarity. This algorithm is called Fourier-domain acceleration searching. This was implemented in CUDA C to accelerate computational-intensive tasks on GPUs. Optimizations were implemented to reduce the computational requirements as well as to perform a more efficient data transfer from host to device. This GPU-accelerated implementation was compared to PRESTO and PRESTO2 ON GPU. Results show that GPU-based implementations of acceleration searching outperform CPU-accelerated implementations when the number of kernel templates or segment size is increased. Despite the tact that this algorithm has a very low compute-to-memory ratio, results show that this GPU-based implementation achieved a peak speed-up of 11 times more than its PRESTO CPU-accelerated counterpart. Moreover, this implementation performed up to 30% better than PRESTO2 ON GPU.