The Many-cOre Technology Investigating Value, Application, deploymenT and Efficiency project or (MOTIVATE for short); is a pathfinder project with the aim of investigating the application of the latest many-core technologies, such as GPUs, to deliver energy and cost efficiencies in the area of radio astronomy High Performance Computing.
Astrophysical radio sources such as Pulsars are excellent probes of extreme physical processes originating from compact sources within our Galaxy and beyond. Generating intermittent radio bursts of milliseconds in duration, as observed here on earth, the signals generated by Pulsars carry valuable information about the physical processes occurring at the source as well as in the intervening interstellar or intergalactic medium. In addition to providing information about the distribution of the interstellar medium within our galaxy the intrinsic shape and polarisation properties of the radio pulses received reflects the geometry and plasma distribution around the pulsar yielding information on their composition and construction. As radio astronomy enters a new era of sensitivity through the use of next generation radio telescopes such as the proposed Square Kilometre Array, the ability to probe the physics of these “astrophysical transients” at meaningful resolutions is becoming a reality.
However, although the SKA can detect these radio bursts there is an added complication. As radio signals move from the star to the detector they experience a phenomenon called dispersion as a by-product of the distance travelled through the interstellar medium. A similar effect can be observed when a pulse of light passes through a fibre optic, the pulse takes longer to get out of the fibre than it does to go in. The effect of this is that the detector array receives different frequencies from the source at slightly different time intervals meaning that the original white light from the source is not seen as it would be if observed from a much shorter distance. As such the signal needs to be de-dispersed to form the original signal.
Although this can be easily achieved if the distance traveled and the nature of the medium is known, the principal aim of SKA and other experiments is to look for new, unknown objects in the universe. As such we don’t know how dispersed any signal will be because we have no idea how far it has traveled and through what parts of the Inter Stellar Medium, hence all potential dispersion values need to be tested. Hence success in detection and classification depends on fast searches across all possible dispersion measures, which requires dedicated high performance computing.
The Oxford team led by Mike Giles and Aris Karastergiou has focused on a “Brute force” de-dispersion algorithm which relies on trying all possible dispersion scenarios to detect a signal from a transient event. This is very computationally intensive but exact, meaning it does not employ approximations and is a true interpretation. Since the algorithm is suited to fine-grained type of parallelism, many-core architectures are an obvious candidate for these types of analyses.
Subsequent refinement of the algorithms and comparison across available GPU and CPU systems revealed that the shared memory GPU algorithm achieved 50% – 70% of the peak GPU performance leaving little margins for improvement for GPU based, brute-force, incoherent de-dispersion algorithms whilst making real-time dispersion searches a possibility under many different observing situations especially in applications such as SKA Phase 1.
As well as boosting the scientific return of these telescopes the excellent performance of GPUs in this application out performs the CPU algorithm in terms of raw performance, cost of hardware and energy providing across the board economic returns as well making GPU clusters such as EMERALD ideal for the task.
Dr. Wes Armour
University of Oxford