Fourier-domain dedispersion

TL;DR

The paper introduces the Fourier-domain dedispersion (FDD) algorithm, a compute-efficient method for correcting dispersion delays in radio signals, outperforming traditional time-domain methods especially for large numbers of dispersion measures.

Contribution

It presents the FDD algorithm as a GPU-optimized, compute-limited alternative to traditional time-domain dedispersion, enabling faster processing of radio astronomical data.

Findings

FDD outperforms optimized time-domain dedispersion by 20% in speed.

FDD reduces energy consumption by 5% compared to traditional methods.

Performance gains increase with larger numbers of dispersion measures (>=512).

Abstract

We present and implement the concept of the Fourier-domain dedispersion (FDD) algorithm, a brute-force incoherent dedispersion algorithm. This algorithm corrects the frequency-dependent dispersion delays in the arrival time of radio emission from sources such as radio pulsars and fast radio bursts. Where traditional time-domain dedispersion algorithms correct time delays using time shifts, the FDD algorithm performs these shifts by applying phase rotations to the Fourier-transformed time-series data. Incoherent dedispersion to many trial dispersion measures (DMs) is compute, memory-bandwidth and I/O intensive and dedispersion algorithms have been implemented on Graphics Processing Units (GPUs) to achieve high computational performance. However, time-domain dedispersion algorithms have low arithmetic intensity and are therefore often memory-bandwidth limited. The FDD algorithm avoids this limitation and is compute limited, providing a path to exploit the potential of current and upcoming generations of GPUs. We implement the FDD algorithm as an extension of the DEDISP time-domain dedispersion software. We compare the performance and energy-to-completion of the FDD implementation using an NVIDIA Titan RTX GPU against the standard as well as an optimized version of DEDISP. The optimized implementation already provides a factor of 1.5 to 2 speedup at only 66% of the energy utilization compared to the original algorithm. We find that the FDD algorithm outperforms the optimized time-domain dedispersion algorithm by another 20% in performance and 5% in energy-to-completion when a large number of DMs (>=512) are required. The FDD algorithm provides additional performance improvements for FFT-based periodicity surveys of radio pulsars, as the FFT back to the time domain can be omitted. We expect that this computational performance gain will further improve in the future.