|Submitter:||Hanno Spreeuw, Ben van Werkhoven and Sarod Yatawatta|
|Description:|| SAGECal [1, 2] has been designed to meet not only the challenges of calibrating present-day radio telescopes like LOFAR, but also to handle the data rates of telescopes under construction, like Square Kilometre Array (SKA). To validate this, we have developed new GPU accelerated code and measured the times Sagecal needs to predict the sky and beam for five artificial data sets, having 64, 128, 256, 384 and 512 stations and for five numbers of sources. The size of these datasets obviously increases with the square of the number of stations (N) since the number of baselines equals 0.5xN(N-1). In this figure we depict sky plus beam prediction times for both the CPU as well as the GPU version of Sagecal. For this research, we have used a cluster node equipped with two CPUs (2x Xeon E5-2660v3, 40 logical cores) and a Titan-X (Pascal) GPU on ASTRON's DAS5.|
Both versions depend quadratically on the number of stations, but the GPU version is about ten times faster. In this research, we have explored GPU utilisation and latencies to improve Sagecal performance. We have reduced latencies in the two most important kernels. Our future work involves optimization of a number of other kernels within Sagecal, including the kernels that do the actual calibration.
 S. Yatawatta, S. Kazemi, and S. Zaroubi. GPU accelerated nonlinear optimization in radio interfer-
ometric calibration. In 2012 Innovative Parallel Computing (InPar), pages 1–6, May 2012.