The Performance of Gridding/Degridding on the Cell/B.E.
A.L. Varbanescu
A.S. van Amesfoort
H.J. Sips (TU Delft - Electrical Engineering, Mathematics and Computer Science)
T Cornwell
B. Elmegreen
A Mattingly
G van Diepen
Rob van Nieuwpoort
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
In this report we present our experience with porting two very time consuming kernels from radioastronomy imaging, i.e. the gridding/degridding operations. Both these procedures are implemented using convolutional resampling, a bandwidth-limited application which is not trivial to parallelize on a hybrid memory architecture. We briefly discuss the role of the application in the radioastronomy context, and we present the Cell/B.E. multi-core processor as an interesting target for this application. Further, we show how the original reference implementation (sequential, C++ code) was not “Cell-friendly”, which is also why our first parallel version had poor results. Next, we show how the reference algorithm was tuned to become more suitable for Cell parallelization. We discuss how we have improved our parallelization strategy and we present our performance results with the new tuned algorithm running on the Cell/B.E. processor, showing a speed-up factor of over 20 when compared with the reference implementation. We conclude that a single Cell/B.E. can provide good results for a small-scale version of the radioastronomy imaging problem, and we sketch our future work directions towards using a parallel machine with multiple Cell/B.E. processors to address the real scale of the of the gridding/degridding applications.