The Square Kilometre Array (SKA), the next-generation radio telescope has Exascale compute requirements.


On top, it requires petabyte/s bandwidth of 4-5 orders of magnitude higher than current radio telescopes. Within DEEP-ER we work with key algorithms used by radio telescopes and adapt them to assess state-of-the-art accelerator and network technologies. Two main applications are ported to the DEEP-ER prototype: a “correlator” and an “imager”.

The correlator

  • In a first step, the "correlator" – an application that combines data from individual telescopes – is implemented. This application includes signal-processing algorithms like FIR filters, FFTs, correlations. The performance of the Intel® Xeon PhiTM and the amount of effort needed to optimise the code for this processor are assessed.
  • Additionally, an all-to-all data exchange of the input data over a fast interconnect is necessary to combine the samples from all telescopes. The bandwidth and switching capabilities of the DEEP-ER interconnect are studied in this context. Currently the correlator handles up to 240 Gb/s data and produces up to 80 Gb/s output. In 2016, for the phase 1 of SKA, it must scale to more than 100 Tb/s. The computational requirements grow to more than 1016 operations per second, also three orders of magnitude more than for LOFAR. Another two orders of magnitude are needed in 2020 for SKA phase 2.
  • The biggest I/O challenge is to reorder all input data across the entire system (as each input contains all frequencies from a single receiver, while a processor needs a single frequency from all receivers, to correlate the data), in a streaming, real-time way.

Hence, this application is a good demonstrator for the DEEP-ER interconnect, while it allows us to study scalability beyond current-generation radio telescopes. Also, similar all-to-all exchanges are necessary for other astronomical data processing pipelines.

The imager

In a second stage the “imager'' application, used to create sky images from correlated data, is ported. This application is computationally expensive, since it requires the computation of many convolutions. Recently a new parallel algorithm that maps very well to GPUs has been devised.

Key developments within DEEP-ER

The two key operations are "gridding" of the calibrated correlator output and 2D FFTs to create sky images. The gridding operation involves convolving the correlated data, albeit in a less regularly structured way than classical image convolution algorithms. The necessary computational power scales roughly with the fourth power of the distance between the two outermost telescopes (and thus the resolution of a sky image), and rapidly increases to exa-scale proportions for the SKA phase 2.