SPIE Membership Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
SPIE Photonics West 2019 | Register Today

SPIE Defense + Commercial Sensing 2019 | Call for Papers

2019 SPIE Optics + Photonics | Call for Papers



Print PageEmail PageView PDF


Heterogeneous computing can boost radio astronomy research

Diverse platforms provide relatively low-cost computing power and enable advances in computationally intensive science.
9 February 2011, SPIE Newsroom. DOI: 10.1117/2.1201101.003486

Modern architectures suited for general-purpose computing are often not the best choice for either input/output or compute-bound problems, both common in science. Many trade-offs must be made in a general-purpose computer design, such as between performance and cost or latency and throughput. Heterogeneous systems avoid these compromises by combining computer subsystems designed for specific purposes with a general-purpose computer, resulting in a system that takes advantage of each subsystem's best features.1

A heterogeneous computing system comprises several architecturally diverse, user-programmable processing units. For example, a modern personal computer with a graphical processing unit (GPU) is a heterogeneous computing platform because the user can program its central processing unit (CPU) as well as the GPU to accomplish the task at hand.

We are using heterogeneous computing platforms to build radio astronomy instruments. In such diverse systems, the CPU and GPU are fed by front-end hardware composed of massively parallel, reconfigurable computing elements (RCEs) built with field-programmable gate arrays (FPGAs). These custom RCEs can accommodate the very high data rates present at the input of a radio telescope's signal-processing system. The RCEs' main function is to pre-process the data and either reduce its volume or reorder and repackage it to allow multiple CPU- or GPU-based systems to be used for the calculations.

Radio telescopes do not directly output images. Instead, they capture data from radio-frequency signals that can be manipulated to form images. In many radio astronomy experiments, the data are never used to make sky images, but instead they are employed to explore other dimensions of the signals, such as spectral features or time variability of point sources. A case in point is pulsar science.

Pulsars, rapidly rotating neutron stars, were discovered in 1967 by Jocelyn Bell-Burnell and Antony Hewish at the University of Cambridge, UK. Pulsar astronomy2 studies the physics of these neutron stars and their companions. Observations aim to find new pulsars and to time their spin periods and pulse arrival times. The results of these search and timing observations are used to study neutron stars and research their interaction with black holes. A consortium of astronomers, members of the North American Nanohertz Observatory for Gravitational Waves (NANOGrav),3 are using very accurate pulsar arrival times to search for evidence of gravity waves.

In the field of single-dish radio astronomy, pulsar observers have driven the advancement of digital signal processing. This is because of the insatiable appetite for bandwidth and storage for pulsar searches as well as the need for extreme amounts of computing power for high-precision timing studies. These studies eliminate the dispersive effects of the interstellar medium on the pulsar signals using a technique called coherent dedispersion.2 Interstellar dispersion is analogous to the dispersion encountered when launching a light pulse down an optical fiber. Higher frequencies travel through the fiber at a different speed than lower ones, thereby smearing the pulse in time. This dispersive effect is removed from pulsar signals by applying a filter consisting of an inverse transfer function of the interstellar medium, allowing recovery of a much sharper signal. This process must be applied to the raw data in real time, as the data rates are too high to store the raw samples. In addition, the data rates cannot be reduced by averaging (without losing the information needed to accomplish the dedispersion).

The Green Bank Ultimate Pulsar Processing Instrument (GUPPI)1 uses a heterogeneous computing approach to provide the compute power needed to dedisperse pulsar data in real time at an affordable cost. The GUPPI design derives in part from the very successful Green Bank Astronomical Signal Processor (GASP), which uses a cluster of 20 standard computers to perform the dedispersion on just 64MHz of bandwidth. In contrast, GUPPI dedisperses 800MHz of bandwidth using only eight heterogeneous computers consisting of a CPU and a GPU. Both machines use an FPGA as a front-end processor to collect and distribute data to the cluster.

Figure 1 shows the GASP bandpass plotted on top of the GUPPI bandpass and indicates the tremendous increase in bandwidth available with GUPPI. There are many effects in the wider bandpass that are not evident from using the much narrower GASP, including scintillation effects in parts of the band.

Figure 1. The Green Bank Ultimate Pulsar Processing Instrument (GUPPI) and the Green Bank Astronomical Signal Processor (GASP) overlaid bandpasses demonstrate the improved capabilities of GUPPI. (PSR J1713+0747 plot by S. Ransom, National Radio Astronomy Observatory, or NRAO). BW: Bandwidth. Scintle: Interstellar scintillation. PSR J1713+0747: A binary radio pulsar.

Figure 2 shows the improved pulse profile provided by the signal's coherent dedispersion. The notch on the right side of the pulse profile is completely lost using a simpler incoherent system. High-resolution pulse profiles are necessary to accurately determine the pulse time of arrival. For programs like NANOGrav to succeed, they will need wideband coherent dedispersion systems like GUPPI attached to their telescopes.

Figure 2. Pulse profile improvement with coherent dedispersion (plot by P. Demorest, NRAO). PSR B1937 + 21: Fastest known rotating pulsar. GBT: Green Bank Telescope.

Heterogeneous computing provides an opportunity for radio astronomers to vastly increase the computing power applied to their research. Studies that could not have been done in the past are now routine. The future holds even more promise for heterogeneous computing as the major manufacturers of computing hardware, such as AMD, Intel, and NVIDIA, all introduce products with heterogeneous architectures. We plan to continue building instruments using heterogeneous computing platforms to reduce cost and quicken our time to science output from new instruments.

The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities Inc.

John Ford
National Radio Astronomy Observatory
Green Bank, WV

John Ford, electronics division head, focuses on computer architecture, signal processing, and pulsar research instrumentation and applications.

1. J. M. Ford, P. Demorest, S. Ransom, Heterogeneous real-time computing in radio astronomy, Proc. SPIE 7740, pp. 77400A, 2010. doi:10.1117/12.857666  
2. D. Lorimer, M. Kramer, Handbook of Pulsar Astronomy, Cambridge University Press, Cambridge, UK, 2005.
3. . North American Nanohertz Observatory for Gravitational Waves, http://www.nanograv.org Accessed 28 December 2010.