Speaker
Dr
Sebastiaan van der Tol
Description
In radio astronomical interferometry and other applications, the
measurements and the image are related by an (approximate) Fourier
transform.
In these cases, it is often necessary to resample the measurements onto a
regular grid to be able to use the Fast Fourier Transform (FFT).
Resampling includes a convolution to suppress aliasing. The convolution
function can also include a correction for deviations of the measurement
equation from a Fourier transforms, for example instrumental or atmospheric
effects.
Especially for high update rates of the correction, this can become
computationally costly.
For LOFAR (and future radio observatories) the data volumes are too large to
be send to the end user for further processing.
The data needs to processed at LOFAR central processing.
The processing pipeline needs to run near real time, otherwise an ever
growing backlog will arise.
This requirement could not be met when quickly varying corrections for
atmospheric effects where included using the conventional approach.
Image Domain Gridding (IDG) is a convolutional resampling algorithm designed
from the start to maximize parallelism.
The result is an algorithm that is not the most computationally efficient in
pure operation count, but maps very well onto massively parallel
architectures. It outperforms other approaches that do fewer compute
operations, but are not optimized for parallelism.
Within the DOME project this algorithm has been implemented, optimized and
benchmarked for various parallel architectures.
Within the OBELICS project we have analyzed the accuracy of the algorithm,
embedded it into an imager for the LOFAR pipeline, and benchmarked the
overall performance. Demonstrating that the LOFAR requirements can be met
using the GPUs that are part of the LOFAR cluster
