Reliable and energy-efficient next-generation data centres

数据中心已经变成了主要的资源消耗者，每年的运营预算高达数百万美元。这项工作的目标是通过解决两个主要因素来更有效地运行数据中心：第一个是提供系统可靠性的挑战。30%-40%的运营预算由系统故障和防止系统故障的努力费用构成。第二个是能源使用。一个典型的数据中心超过三分之一的运营预算用于电力消耗。这两个问题正在成为数据中心设计的限制因素，预计未来随着数据中心规模的不断扩大，情况还会变得更糟。**虽然人们对这些问题进行了大量研究，但它们在很大程度上仍未得到解决，主要是因为传统的计算机系统研究方法(实验、模拟、分析)正达到极限：由于复杂性和规模，在实验室试验台上识别和复制真实世界大型系统中的问题极其困难。对数据中心中数以百万计的单个组件如何运行和交互的了解有限，因此很难构建真实的模型或模拟器。我在大规模系统设计中解决问题的方法是，首先从生产机器收集和分析大量现场数据，以详细了解问题的性质，然后使用这些数据来推动解决问题的分析和实验工作。**该提案的长期愿景是更可靠、更节能、更易于管理的数据中心。为了实现这一愿景，拟议的工作将侧重于我在最近的工作中确定的四个具体的近期目标，这些目标对于消除当今数据中心效率低下的主要根源至关重要。从学术角度来看，我们的研究将为大规模系统如何在该领域运行创造更好的理解和更现实的模型，为在一个目前过于依赖假设和轶事的重要领域进行严格研究提供基础。

Datacenters have turned into major resource hogs with annual operating budgets running into many millions of dollars. The goal of this work is to run datacenters more efficiently by addressing two of the main contributing factors: The first is the challenge of providing system reliability. Between 30--40\% of the operating budget is made up by the cost of system failures and efforts to protect against them. The second is energy usage. More than a third of the operating budget of a typical datacenter goes into electricity consumption. Both problems are becoming limiting factors in data center design and are predicted to get worse in the future as data centers keep growing in size.**While much research has been devoted to these problems, they remain largely unsolved, mostly because traditional approaches to computer systems research (experiments, simulation, analysis) are reaching their limits: Due to the complexity and scale it is extremely difficult to identify and replicate problems in real-world large-scale systems in experimental lab testbeds. The limited understanding of how the millions of individual components in data centers operate and interact makes it very hard to build realistic models or simulators. My approach to addressing problems in large-scale system design is to first collect and analyze large amounts of field data from production machines to derive a detailed understanding of the nature of a problem and then use the data to drive the analytical and experimental work for solving a problem.**The long-term vision of this proposal are data centers that are more reliable, more energy-efficient and easier to manage. To achieve this vision the proposed work will focus on four specific near-term goals that I identified in recent work as crucial to combat major sources of inefficiencies in today's data centres. From an academic point of view, our research will create a better understanding and more realistic models of how large-scale systems operate in the field, providing a foundation for rigorous research in an important area that is currently relying too much on assumptions and anecdotes.**********************