NeTS:Small:Understanding the Impact of Unreliable Hardware on the Resilience of Networked Systems

网络系统一直被设计为即使在存在故障的情况下也能运行，特别是在通信链路和存储方面。直到最近，这类系统的其他组成部分的故障概率相对较低，对于大多数联网系统，可以通过使用以特别方式添加的最小冗余来实现所需的复原力水平。两种相反的趋势可能会使未来几年实现弹性的任务变得更加困难：（a）硬件故障概率增加：随着向更精细的纳米级制造迈进，芯片越来越容易受到外部噪声引起的软错误的影响，并且越来越有可能因疲劳而过早失效;（B）更高的弹性要求：随着关键服务不断迁移到云，服务提供商被迫签订更严格的服务级别协议（SLA），包括更高的可靠性、更高的可用性和更严格的服务时间保证。上述组合可能会大大增加现有方法的开销，以实现所需的弹性水平。智力价值：该项目的第一个成果将是网络系统弹性的整体路线图。该弹性路线图将采用纳米级CMOS的路线图（芯片成本，功能，性能，功耗和可在芯片级实现的弹性的趋势），并试图实际预测当前使用的网络和系统技术的未来成本，以实现所需的弹性水平。该项目的第二个成果是开发弹性方法，这些方法可以在面对越来越多的硬件故障时优雅地扩展。此类技术将使用新型分区冗余策略，在硬件和软件层的不同级别实现可靠性。更广泛的影响。复原力路线图将提供对复原力趋势的前所未有的理解，以及对挑战和机遇的独特现实评估。这将对硬件和网络社区的研究产生重大影响。可扩展的弹性方法的系统设计将导致显著更高水平的弹性，更低的成本-资本（设备）以及经常性（特别是能源），和/或更高水平的性能。 由于网络系统现在已成为我们最重要的基础设施之一，消耗的资源比例越来越大，因此拟议项目对社会的实用收益可能是巨大的。该项目将借鉴两个不同的学科，硬件架构和网络系统，并涉及详细的案例研究和全新理论和技术的发展，预算影响说明：本段的项目编号是指我们原来建议的图9和第3.2节（题为“建议的研究工作及计划”）所载的项目编号。我们将承担第1项（及其所有分项）中提出的所有任务和分任务。在第2项中，我们将着手制定一个总体框架，以考虑所有基本的冗余计划和部署这些计划的替代方法（分项2.1）。我们还将描述相关的权衡（分项2.2）和现实约束的后果（分项2.3）。然而，我们将继续开发原型工具（如分项2.4所述），以证明我们的方法的好处并进行案例研究（如第3项所述）。最后，我们将进行项目3中最初提议的案例研究。

Networked systems have always been designed to operate even in the presence of failures, especially in communication links and storage. Until recently other components of such systems had relatively low probabilities of failures and for most networked systems, desired levels of resilience could be achieved using minimal redundancy added in an ad hoc manner. Two opposing trends are likely to make the task of achieving resilience significantly more difficult in the coming years: (a) increasing hardware failure probabilities: with the move towards finer nano-scale fabrication, chips are increasingly vulnerable to soft errors caused by external noise and are increasingly likely to fail early due to fatigue; (b) higher resilience requirements: as critical services continue to migrate to clouds, service providers are compelled into more stringent service-level agreements (SLAs), including higher reliability, higher availability, and tighter guarantees on service times. The above combination can dramatically increase the overhead of existing approaches for achieving desired levels of resilience. Intellectual merit: The first outcome of this project will be a holistic roadmap for resilience of networked systems. This resilience roadmap will take the roadmaps from the nano-scale CMOS (trends in chip cost, functionality, performance, power, and resilience that can be attained at chip level) and attempt to realistically project the future cost of currently-used networking and systems techniques for achieving desired level of resilience. The second outcome of this project is to develop resilience methods that scale gracefully in the face of increasing hardware failures. Such techniques will use novel partitioned redundancy strategies that achieve reliability at different levels across hardware and software layers. Broader Impacts. The resilience roadmap will provide unprecedented understanding of the trends in resilience and a uniquely realistic assessment of challenges and opportunities. This will significantly influence the research in the hardware as well as networking communities. A systematic design of scalable resilience methods will lead to significantly higher levels of resilience, lower costs - capital (equipment) as well recurring (especially, energy), and/or higher levels of performance. The utilitarian gains to society by the proposed project are likely to be substantial, since networked systems now constitute one of our most critical infrastructures and consume an increasingly large proportion of our resources.This project will draw upon two different disciplines, hardware architecture and networked systems, and involve detailed case studies and development of completely new theory and techniques, and will therefore provide unique educational and training opportunities for students and working professionals in these fields.Budget Impact Statement: The item numbers in this paragraph refer to those in Figure 9 and Section 3.2 (entitled 'Proposed Research Tasks and Plan') of our original proposal. We will undertake all tasks and sub-tasks proposed in item-1 (and all its sub-items). In item-2, we will undertake the development of a general framework to consider all basic redundancy schemes and alternative ways of deploying them (sub-item-2.1). We will also characterize the associated tradeoffs (sub-item-2.2) and the consequences of realistic constraints (sub-item-2.3). However, we will pursue the development of prototype tools (as outlined in sub-item-2.4), to the extent necessary to demonstrate the benefits of our approach and to conduct case studies (described in item-3). Finally, we will undertake the case studies as originally proposed in item-3.