Reducing adaptive optics latency using many-core processors
Abstract
Atmospheric turbulence reduces the achievable resolution of ground based optical
telescopes. Adaptive optics systems attempt to mitigate the impact of this turbulence
and are required to update their corrections quickly and deterministically (i.e. in realtime).
The technological challenges faced by the future extremely large telescopes
(ELTs) and their associated instruments are considerable. A simple extrapolation of
current systems to the ELT scale is not sufficient.
My thesis work consisted in the identification and examination of new many-core
technologies for accelerating the adaptive optics real-time control loop. I investigated
the Mellanox TILE-Gx36 and the Intel Xeon Phi (5110p). The TILE-Gx36 with
4x10 GbE ports and 36 processing cores is a good candidate for fast computation of
the wavefront sensor images. The Intel Xeon Phi with 60 processing cores and high
memory bandwidth is particularly well suited for the acceleration of the wavefront
reconstruction.
Through extensive testing I have shown that the TILE-Gx can provide the performance
required for the wavefront processing units of the ELT first light instruments.
The Intel Xeon Phi (Knights Corner) while providing good overall performance does
not have the required determinism. We believe that the next generation of Xeon Phi
(Knights Landing) will provide the necessary determinism and increased performance.
In this thesis, we show that by using currently available novel many-core processors
it is possible to reach the performance required for ELT instruments.