Accelerating radio transient detection using the Bispectrum algorithm and GPGPU

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

Modern radio interferometers such as those in the Square Kilometre Array (SKA) project are powerful tools to discover completely new classes of astronomical phenomena. Amongst these phenomena are radio transients. Transients are bursts of electromagnetic radiation and is an exciting area of research as localizing pulsars (transient emitters) allow physicists to test and formulate theories on strong gravitational forces. Current methods for detecting transients requires an image of the sky to be produced at every time step. Since interferometers have more information available to them, the computational demands for producing images becomes infeasible due to the larger data sets provided by larger interferometers. Law and Bower (2012) formulated a different approach by using a closure quantity known as the "bispectrum": the product of visibilities around a closed loop of antennae. The proposed algorithm has been shown to be easily parallelized and suitable for Graphics processing units (GPUs).Recent advancements in the field of many core technology such as GPUs has demonstrated significant performance enhancements to many scientific applications. A GPU implementation of the bispectrum algorithm has yet to be explored. In this thesis, we present a number of modified implementations of the bispectrum algorithm, allowing both instruction-level and data-level parallelism. Firstly, a multi-threaded CPU version is developed in C++ using OpenMP and then compared to a GPU version developed using Compute Unified Device Architecture (CUDA).In order to verify validity of the implementations presented, the implementations were firstly run on simulated data created from MeqTrees: a tool for simulating transients developed by the SKA. Thereafter, data from the Karl Jansky Very Large Array (JVLA) containing the B0355+54pulsar was used to test the implementation on real data. This research concludes that the bispectrum algorithm is well suited for both CPU and GPU implementations as we achieved a 3.2x speed up on a 4-core multi-threaded CPU implementation over a single thread implementation. The GPU implementation on a GTX670, achieved about a 20 times speed-up over the multi-threaded CPU implementation. These results show that the bispectrum algorithm will open doors to a series of efficient transient surveys suitable for modern data-intensive radio interferometers.