Feng Lu, Fedrigo Enrico, Béchet Clémentine, Brunner Elisabeth, Pirani Werther
National Observatories of China, Datun Road 20A, Beijing 100012, China.
Appl Opt. 2012 Jun 1;51(16):3564-83. doi: 10.1364/AO.51.003564.
The European Southern Observatory (ESO) is studying the next generation giant telescope, called the European Extremely Large Telescope (E-ELT). With a 42 m diameter primary mirror, it is a significant step from currently existing telescopes. Therefore, the E-ELT with its instruments poses new challenges in terms of cost and computational complexity for the control system, including its adaptive optics (AO). Since the conventional matrix-vector multiplication (MVM) method successfully used so far for AO wavefront reconstruction cannot be efficiently scaled to the size of the AO systems on the E-ELT, faster algorithms are needed. Among those recently developed wavefront reconstruction algorithms, three are studied in this paper from the point of view of design, implementation, and absolute speed on three multicore multi-CPU platforms. We focus on a single-conjugate AO system for the E-ELT. The algorithms are the MVM, the Fourier transform reconstructor (FTR), and the fractal iterative method (FRiM). This study enhances the scaling of these algorithms with an increasing number of CPUs involved in the computation. We discuss implementation strategies, depending on various CPU architecture constraints, and we present the first quantitative execution times so far at the E-ELT scale. MVM suffers from a large computational burden, making the current computing platform undersized to reach timings short enough for AO wavefront reconstruction. In our study, the FTR provides currently the fastest reconstruction. FRiM is a recently developed algorithm, and several strategies are investigated and presented here in order to implement it for real-time AO wavefront reconstruction, and to optimize its execution time. The difficulty to parallelize the algorithm in such architecture is enhanced. We also show that FRiM can provide interesting scalability using a sparse matrix approach.
欧洲南方天文台（ESO）正在研究下一代巨型望远镜，即欧洲极大望远镜（E-ELT）。其主镜直径达42米，相较于现有望远镜有了重大提升。因此，E-ELT及其仪器在成本和控制系统的计算复杂性方面带来了新挑战，包括其自适应光学（AO）系统。由于目前在AO波前重建中成功使用的传统矩阵向量乘法（MVM）方法无法有效地扩展到E-ELT上AO系统的规模，所以需要更快的算法。在最近开发的那些波前重建算法中，本文从设计、实现以及在三个多核多CPU平台上的绝对速度等角度研究了三种算法。我们聚焦于E-ELT的单共轭AO系统。这些算法分别是MVM、傅里叶变换重建器（FTR）和分形迭代方法（FRiM）。这项研究提高了这些算法在参与计算的CPU数量增加时的扩展性。我们根据各种CPU架构限制讨论了实现策略，并给出了到目前为止在E-ELT规模下的首个定量执行时间。MVM存在巨大的计算负担，使得当前的计算平台规模不足以达到AO波前重建所需的足够短的时间。在我们的研究中，FTR目前提供最快的重建速度。FRiM是一种最近开发的算法，本文研究并提出了几种策略，以便将其用于实时AO波前重建并优化其执行时间。在这种架构中并行化该算法的难度增加。我们还表明，使用稀疏矩阵方法，FRiM可以提供有趣的扩展性。
