Oral Abstract

Oral Contribution (O9.4) Mateusz Malenta (The University of Manchester)

Theme: Data processing pipelines

Accelerating Radio Astronomy With High Performance Computing

In my talk, I will present the challenges that real-time radio transient surveys face from inside one of the projects currently scanning the sky. MeerTRAP is a commensal, real-time pulsar and fast transients survey that will be conducted using MeerKAT - an interferometric precursor to the Square Kilometre Array, consisting of 64 dishes in South Africa.
The main goal of this project is to detect and localise fast radio transients, including FRBs, RRATs and pulsars as close to real time as possible. Real-time detection is especially important when dealing with millisecond-duration bursts so that follow-up observations at different wavelengths can be started. This requires search pipelines to be capable of both rapidly processing large amounts of data, of the order of hundreds of gigabits per second, and presenting the observer with output that can be easily analysed with as little delay as possible.
I will discuss the off-the-shelf and custom-developed HPC solutions that form the backbone of modern real-time pipelines, with an emphasis on GPU computing and NUMA-aware multithreaded applications and their associated challenges. I will also outline the roadmap for future developments, including current and future requirements with regards to faster and more efficient computing, and also reducing the need for human intervention with the use of Machine Learning.
On a higher level, I will discuss the challenges of managing tens of clusters, and solutions offered by modern containerisation and orchestration technologies. Finally, I will summarise the experiences from this project which can be used to advise the radio astronomy community.