Birds of a Feather Discussion (B5) Yan Grange (ASTRON, the Netherlands Institute for Radio Astronomy)
Accelerating scientific discoveries in the modern hardware landscapeThis session will start with a few speakers after which a discussion will be kicked off:
1) Yan Grange & Raymond Oonk: short introduction, setting the stage
2) Sagar Dolas (SURFsara) on Sustainable GPU programming practices
3) Andre Gunst (ASTRON) on ASTRONs experiences with FPGAs
Astronomy has been a field with computational challenges. Data sets grow larger and larger, and the number of computations per unit data widely varies, making that traditional HPC is not always well-fit. This calls for development of algorithms that make use of hardware
accelerators, like GPUs, FPGAs, ASICs, Quantum computers, or even hybrid solutions. \r\nHardware accelerators have been used as means to increase processing capacity for specific use cases but the massive adoption of algorithms built for those lags behind when considering the user processing use cases (like calibration and imaging of radio-interferometric data).
This is partly due to the vicious circle that production compute facilities (both observatory and private clusters) do not contain accelerators due to a lack of algorithms while algorithms do not get the chance to mature in operational readiness. Also, ensuring reproducibility requires code or binaries that are as generic as possible while optimisation for specific hardware platforms can greatly improve performance.
The goal of the session is to discuss how to move the development forward, taking into account the platforms. The discussion will mainly focus around questions, such as:
(i) How to integrate HW acceleration of specific tasks in a complex workflow?
(ii) How to make workflows both optimised and portable?
(iii) What is the role of data location and transport, etc.?
(iv) How to bring accelerated code from research to operations?\r\n\r\nBy solving these questions, we can more efficiently leverage high-performance accelerators, to advance astronomy and accelerate scientific discoveries.Sign on url: