ITR/AP Methodologies and Tools for Designing and Implementing Large Scale Real-Time Systems

该提案要求支持开发用于设计和实现超大规模实时嵌入式计算机系统的方法和工具的研究，所述方法和工具(A)通过使用并行硬件体系结构来实现超高计算性能，(B)通过分布式、分级监视和控制来实现和维护功能完整性，(C)要求高度可用，以及(D)是动态可重构、可维护和可进化的。推动这项研究并为其提供测试平台的具体应用是BTeV的触发和数据采集系统，BTeV是一项基于加速器的高能物理(HEP)实验，用于研究含有b夸克的粒子衰变中的物质-反物质不对称(也称为电荷宇称破坏)。BTeV已获得费米实验室管理层的批准，预计将在未来5-6年内与费米实验室Tevatron对撞机一起运行。该实验的数据采集阶段预计至少需要五年时间。它需要一个大规模并行的、不同种类的计算元素集群来重建每秒1500万个粒子相互作用(事件)，并使用重建数据来决定保留哪些事件以进行进一步的数据分析。为这种类型的实时嵌入式系统创建可用的软件将需要研究计算机科学和工程领域中的一般问题的解决方案。倡导者计划以一种普遍的方式处理这些问题，并制定出可应用于许多科学和商业问题的方法和工具。在这个项目期间，研究成果将通过项目进入高中系统，这些项目将加强现有的基础设施，以吸引学生进入科学和工程学科。这项研究的目标系统类别包括那些嵌入在BTeV等环境中的系统，这些环境产生必须使用依赖于数据的计算策略实时处理的非常大的数据流。这些系统需要超高性能(约1012次/秒)，这就需要并行硬件架构，就BTeV而言，该架构由数千个商用处理器、诸如数字信号处理器(DSP)等专用处理器以及诸如现场可编程门阵列(FGA)等专用硬件组成，所有这些都通过非常高速的网络连接。系统必须是可动态重新配置的，以实现可用资源和潜在变化资源的最高性能。系统必须高度可用，因为环境在很长一段时间内持续产生数据流，而且对分析很重要的有趣现象很少见，并且随时可能发生。为了实现高可用性，系统必须具有容错、自我感知和故障自适应能力，因为处理元件、互连开关或前端传感器(提供输入流)的任何故障都可能导致不可恢复的数据丢失。故障必须在尽可能短的时间内纠正，并且必须半自动地纠正(即，尽可能少的人为干预)。因此，分布式和层次化的监测和控制至关重要。

This proposal requests support for research to develop methodologies and tools for designing and implementing very large-scale real-time embedded computer systems that (a) achieve ultra high computational performance through use of parallel hardware architectures, (b) achieve and maintain functional integrity via distributed, hierarchical monitoring and control, (c) are required to be highly available, and (d) are dynamically reconfigurable, maintainable, and evolvable. The specific application that will drive this research and provide a test platform for it is the trigger and data acquisition system for BTeV, an accelerator-based High Energy Physics (HEP) experiment to study matter-antimatter asymmetries (also known as Charge-Parity violation) in the decays of particles containing the b-quark. BTeV has been approved by Fermilab management and is expected to be constructed over the next 5-6 years to run in conjunction with the Fermilab Tevatron Collider. The data-taking phase of the experiment is expected to be at least five years. It requires a massively parallel, heterogeneous cluster of computing elements to reconstruct 15 million particle interactions (events) per second and to use the reconstruction data to decide which events to retain for further data analysis. Creating usable software for this type of real-time embedded system will require research into solutions of general problems in the fields of computer science and engineering. The proponents plan to approach these problems in a way that is general, and to produce methodologies and tools that can be applied to many scientific and commercial problems. During this project, the research results will be carried into the high-school system through projects, which enhance existing infrastructure for attracting students into science and engineering disciplines. The classes of systems targeted by this research include those imbedded in environments, like BTeV, that produce very large data streams which must be processed in real-time using data-dependent computation strategies. These systems require ultra high performance (~1012 operations per second), necessitating parallel hardware architectures, which in the case of BTeV is composed of a mix of thousands of commodity processors, special purpose processors such as Digital Signal Processors (DSPs), and specialized hardware such as Field Programmable Gate Arrays (FPGAs), all connected by very high-speed networks. The systems must be dynamically reconfigurable to allow maximum performance from the available and potentially changing resources. The systems must be highly available, since the environments produce the data streams continuously over a long period of time, and interesting phenomena important to the analysis are rare and could occur at any time. To achieve the high availability, the systems must be fault tolerant, self-aware, and fault adaptive, since any malfunction of processing elements, the interconnection switches, or the front-end sensors (which provide the input stream) can result in an unrecoverable loss of data. Faults must be corrected in the shortest possible time, and corrected semi-autonomously (i.e., with as little human intervention as possible). Hence, distributed and hierarchical monitoring and control are vital.