SAGECal: Full acceleration ahead
Computation of the model involves evaluation of many millions of transcendental functions that are expensive CPU operations. Using GPU acceleration, we get an order of magnitude speedup as shown in Figure 1. We have also shown the scalability of the
To look deeper in the Universe, radio telescopes are becoming more and more sensitive, at the cost of delivering larger and larger volumes of data. In order to produce images that reveal the faintest sources, we need software to calibrate these large amounts of data in an accurate and efficient manner.
acceleration, expanding from LOFAR to SKA. SAGECal is in good shape to handle increased data volumes generated by any future radio telescope.
Hanno Spreeuw (h.spreeuw@esciencecenter. nl) Sarod Yatawatta (email@example.com) We started developing SAGECal over a decade ago, mainly to serve the LOFAR Epoch of Reionisation key science project. In order to detect weak reionisation signatures from the early Universe, Terabytes of LOFAR data need to be accurately calibrated and calibration has to be done as fast as possible. We have accelerated every operation in SAGECal using Graphics Processing Units (GPU). Initially, the various non-linear optimisation routines used in calibration were accelerated. A major bottleneck thus far has been the computation of the model that is needed to run the optimisation routines. The model consists of thousands of celestial sources, some having exotic components such as shapelets. Additionally, the phased array beam shape needs to be computed along the
Figure 1: SAGEcal sky and beam model computation time for five simulated data sets with stations increasing from 64 to 512, and for five different sky models with varying number of sources. We have used a computer equipped with CPUs (2 Xeon ES-2660-v3, 40 logical cores) and a Titan-X (Pascal) GPU on ASTRONâ€™s DAS5 cluster.
directions of every source.
ASTRON News / Winter 2018