Building a reliable, scalable and affordable RTC for AO instruments on ELTs
Building a reliable, scalable and affordable RTC for AO instruments on ELTs
复制标题
为 ELT 上的 AO 仪器构建可靠、可扩展且经济实惠的 RTC
DOI：
--
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
J. Brulé
中科院分区：
文献类型：
--
作者：
D. Gratadour;A. Sevin;D. Perret;J. Brulé
解决ELT AO仪器实时控制器（RTC）所需的前所未有的计算能力是下一代AO系统设计所需的关键技术发展之一。由于能够进行通用计算的设备的快速出现，诸如GPU的面向CPU的架构（其提供比高端CPU大几个数量级的计算性能）最近已经作为有吸引力的和经济上可行的候选者出现。然而，对于实时应用程序，使用不能由操作系统内部调度或控制但通过闭源驱动程序发送命令的I /O设备会带来许多挑战。基于使用GPU进行几乎实时的端到端仿真的经验，并依赖于COMPASS平台的开发，这是一个用于AO仿真和实时控制的统一优化框架，我们的团队艾德致力于为AO RTC开发一个可扩展的、基于GPU的异构原型。在本文中，我们评估了在实时系统中使用GPU进行AO时所面临的主要挑战，并根据影响重要性和可用解决方案对其进行排名。我们提出了我们的战略，以减轻这些问题，包括我们的原型，实时核心和额外的专用组件的数据采集和分发的一般架构。最后，我们讨论了预期性能的延迟和抖动的基础上现实的基准测试，并专注于尺寸的MICADO AO模块RTC。
Addressing the unprecedented amount of computing power nee ded by the ELTs AO instruments real-time controllers (RTC) is one of the key technolo gical developments required for the design of the next generation AO systems. Throughput oriented arch itectures such as GPUs, providing orders of magnitude greater computational performance than high-en d CPUs, have recently appeared as attractive and economically viable candidates since the fast emergenc e of devices capable of general purpose computing. However, using for real-time applications a I /0 device which cannot be scheduled nor controlled internally by the operating system but is sent commands thro ugh a closed source driver comes with a number of challenges. Building on the experience of almost real-time end-to-end simulations using GPUs, and relying on the development of the COMPASS platform , a unified and optimized framework for AO simulations and real-time control, our team has engag ed into the development of a scalable, heterogeneous GPU-based prototype for an AO RTC. In this paper, we eview the main challenges arising when utilizing GPUs in real-time systems for AO and rank them in terms of impact significance and available solutions. We present our strategy, to mitigate t hese issues including the general architecture of our prototype, the real-time core and additional dedicated components for data acquisition and distribution. Finally, we discuss the expected performance in terms of latency and jitter on the basis of realistic benchmarks and focusing on the dimensioning of the MICADO AO module RTC.