AO4ELT5 Proceedings
A GPU based RTC for E-ELT Adaptive optics : Real Time Controller prototype
Bernard, Julien (LESIA - Observatoire de Paris/ Université Paris Diderot), Gratadour, Damien (LESIA - Observatoire de Paris/ Université Paris Diderot), Lainée, Maxime (LESIA - Observatoire de Paris/ Université Paris Diderot), Perret, Denis (LESIA - Observatoire de Paris/ Université Paris Diderot), Sevin, Arnaud (LESIA - Observatoire de Paris/ Université Paris Diderot)
Most of the control laws used in AO require one or several matrix-vector multiplies, at frequencies around 1 kHz. With the Extremely Large Telescope (ELT) generation, the real-time control of the AO loop becomes one of the most challenging issues due to the high computational power required (large matrices and high frequency) and the limited energy footprint inherent to the telescope location (based in isolated regions). The Green Flash European project is in line with this challenge and aims at building a prototype for a Real-Time Controller (RTC) at the ELT scale. We propose a GPU-based solution because of their energy efficiency and throughput capabilities. In order to meet the requirements of an AO loop, in terms of jitter and throughput, we chose a very low level approach relying on persistent kernels to handle all the computational steps from pixels calibration to slopes and command vectors computation. This approach simplifies the latency management by reducing the communication overheads but led us to re-implement an entire AO control loop relying on some GPUs standard features : communication mechanisms (guard, peer-to-peer), algorithms (generalized matrix-vector multiplication, reduce/all reduce) and new synchronization mechanisms on a multi node-multi GPUs system. In this paper we detail the context and the data pipeline of the RTC and report the performance (time, jitter) we obtained and the scalability of the solutions on the NVidia DGX-1 platform.
DOI: 10.26698/AO4ELT5.0176- Proceeding PDF
