Pulsar Acceleration Searches on the GPU for the Square Kilometre Array

Sofia Dimoudi, Wesley Armour

Abstract: Pulsar acceleration searches are methods for recovering signals from radio telescopes, that may otherwise be lost due to the effect of orbital acceleration in binary systems. The vast amount of data that will be produced by next generation instruments such as the Square Kilometre Array (SKA) necessitates real-time acceleration searches, which in turn requires the use of HPC platforms. We present our implementation of the Fourier Domain Acceleration Search (FDAS) algorithm on Graphics Processor Units (GPUs) in the context of the SKA, as part of the Astro-Accelerate real-time data processing library, currently under development at the Oxford e-Research Centre (OeRC), University of Oxford.

Introduction

Pulsars—highly magnetized, rotating neutron stars—serve as the universe's most precise natural clocks. However, detecting them becomes significantly more complex when they reside in binary systems. The orbital motion of the pulsar creates a Doppler effect: as the pulsar moves toward or away from Earth, its apparent pulse frequency drifts over time.

In a standard Fourier analysis using a Discrete Fourier Transform (DFT), this frequency drift causes the signal's power to "smear" across multiple frequency bins. This smearing often reduces the signal-to-noise ratio to the point where the pulsar becomes invisible against background noise. To recover these signals, astronomers must perform an "acceleration search," which attempts to "de-drift" the signal by testing numerous possible acceleration values (frequency derivatives).

Visualization of normalized Fourier power across frequency and frequency derivative bins, showing the peak signal recovered through acceleration searching.

The upcoming Square Kilometre Array (SKA) presents a massive scaling challenge. It will produce unprecedented data volumes, requiring researchers to search through approximately 6,000 different Dispersion Measures across 8 million samples per beam—all in real-time. Traditional CPU-based methods lack the throughput and power efficiency required for this scale of processing.

Methodology

This research implements the Fourier Domain Acceleration Search (FDAS) algorithm. Unlike time-domain resampling, which requires a computationally expensive full DFT for every acceleration trial, FDAS utilizes a correlation technique in the Fourier domain.

The GPU Pipeline

The pipeline is built using NVIDIA’s CUFFT library for fast Fourier transforms and custom CUDA kernels. The process follows a seven-step sequence:

  1. Data Transfer: Moving time-series data to GPU memory.
  2. Initial DFT: Performing a large real-to-complex DFT on the entire time series.
  3. Segmentation: Dividing the Fourier-domain data into manageable segments.
  4. Complex DFTs: Performing complex-to-complex DFTs on these segments.
  5. Convolution: Applying Finite Impulse Response (FIR) filters that model constant acceleration using the overlap-save method to manage edge effects.
  6. Inverse DFTs: Returning the data to the time domain (or a representation thereof) to identify peaks.
  7. Power Spectrum Calculation: Finalizing the search to identify candidate signals.

Experimental Results

The implementation was tested using synthetic data designed to match SKA specifications: time series of $2^{23}$ samples (approx. 8.4 million samples) representing a 530-second observation window. Performance was evaluated on two NVIDIA architectures: the GeForce Titan X (Maxwell) and the Tesla K80 (Kepler).

Real-Time Performance

The primary benchmark for the SKA is "real-time execution"—the ability to process data at the same rate it is collected. As shown in the performance analysis, both GPUs successfully met this requirement for a significant range of acceleration templates.

Execution time for 6,000 searches as a percentage of real-time vs. the number of acceleration templates for Titan X and K80 GPUs.

The Titan X significantly outperformed the K80, maintaining better-than-real-time performance (under 100% on the Y-axis) for up to approximately 125 acceleration templates. The linear scaling observed in the results suggests that the algorithm's performance is predictable, allowing for precise hardware provisioning as search parameters grow more complex.

Discussion and Future Work

The transition from time-domain to Fourier-domain searching represents a critical shift for radio astronomy. By treating the acceleration search as an "embarrassingly parallel" problem, the GPU can process hundreds of templates simultaneously. Furthermore, the performance-per-watt advantage of GPUs is vital for the SKA, where energy consumption at remote telescope sites is a major operational constraint.

While these preliminary results are promising, future work will focus on integrating FDAS into the full Astro-Accelerate library, which includes other intensive tasks like de-dispersion. Additionally, while the constant acceleration assumption used here is effective for long-period binaries, further optimization may be needed for extremely tight binary systems where acceleration changes rapidly during the observation window. Future iterations will also explore optimizations for newer GPU architectures to further increase the search depth and sensitivity.

Summary updated

Abstract

Pulsar acceleration searches are methods for recovering signals from radio telescopes, that may otherwise be lost due to the effect of orbital acceleration in binary systems. The vast amount of data that will be produced by next generation instruments such as the Square Kilometre Array (SKA) necessitates real-time acceleration searches, which in turn requires the use of HPC platforms. We present our implementation of the Fourier Domain Acceleration Search (FDAS) algorithm on Graphics Processor Units (GPUs) in the context of the SKA, as part of the Astro-Accelerate real-time data processing library, currently under development at the Oxford e-Research Centre (OeRC), University of Oxford.
BibTeX
@inproceedings{dimoudi2015pulsar,
  author = {Sofia Dimoudi and Wesley Armour},
  title = {{Pulsar Acceleration Searches on the GPU for the Square Kilometre Array}},
  booktitle = {Oxford},
  year = {2015},
}
Figure