A data-driven approach to improving data center efficiency

大规模的数据处理和存储已经成为我们现代社会的基础。几乎所有行业部门都依赖于高效的数据管理和存储，作为推动其业务发展的引擎，最终用户越来越依赖于大规模处理和存储信息所提供的服务。因此，将数量惊人的资源（包括货币资源和环境资源）用于数据的存储和管理可能并不奇怪。例如，最近的一项研究得出结论，到2025年，管理“数据海啸”可能会消耗全球五分之一的电力。因此，就未来的经济成功和环境影响而言，优化利用系统来处理和存储大规模数据至关重要。当然，计算集群和数据中心的效率问题并不新鲜。近年来，许多研究都集中在这个问题的许多不同方面，事实上，太多的研究无法在本建议中进行详尽的调查。一些例子包括更节能的硬件和能源比例的服务器，优化数据中心冷却系统以改善气流，提高数据中心温度，新的存储介质，灵活和可扩展的调度器，分解存储以允许单独扩展存储和计算资源，更轻量级的虚拟化机制，持续改进ML和AI算法的效率等等。提出这一建议的关键观察是，即使在今天，我们仍然非常低效地使用数据中心和集群资源。服务器虚拟化和针对高密度进行了优化的超大规模数据中心的出现，以及关于该主题的丰富学术文献给人的印象是，当今的计算基础设施通常得到了有效的利用。相比之下，PI最近与几家运营自己数据中心的大公司的讨论显示，数据中心的利用率仍然大大不足，即使对于超大规模的数据中心，典型的利用率也在5-25%之间。对于较小的系统，利用率甚至更低。本提案中工作的长期目标是识别、描述和消除当今存储和计算基础设施中效率低下的主要来源，并提高其利用率。该方法将遵循PI的标志性研究方法：与行业合作以获取生产机器的真实数据，分析数据以表征和识别问题，并使用严格的统计和算法方法来解决核心问题。该提案概述了PI在其近期工作中确定的一系列具体近期目标，并计划在5年内实现这些目标。她希望研究结果能开辟新的研究思路，为更长期的研究议程提供信息。

Large-scale processing and storage of data has become the underpinning of our modern society. Nearly all industry sectors depend on efficient data management and storage as the engine that moves their business and end-users have grown dependent on services enabled by large-scale processing and storing of information. Consequently, it is maybe not surprising that a staggering amount of resources (both monetary as well as environmental) are dedicated to the storage and management of data. For example, a recent study concludes that managing the `Tsunami of data' could consume one fifth of global electricity by 2025. As a result, making optimal use of systems for processing and storing data at large scale is critical both in terms of future economic success and environmental impact. Of course, the problem of efficiency in compute clusters and data centers is not new. Much research in recent years has focused on many different facets of the problem, too many in fact to include an exhaustive survey in this proposal. Some examples include more energy-efficient hardware and energy-proportional servers, optimizations to the data center cooling system to improve airflow, increasing data center temperatures, new storage media, flexible and scalable schedulers, disaggregated storage to allow separate scaling of storage and compute resources, more light-weight virtualization mechanisms, continuous improvements in the efficiency of ML and AI algorithms and many more. The key observation motivating this proposal is that even today we are still making terribly inefficient use of data center and cluster resources. The advent of server virtualization and hyper-scale data centers, which are optimized for high density, and the rich academic literature on the topic create the impression that today's compute infrastructure is generally efficiently utilized. In contrast, the PI's recent discussions with several large companies operating their own data centers revealed that data centers are still greatly under-utilized, with typical utilization levels in the 5-25% range, even for hyper-scale data centers. For smaller systems utilization is believed to be even worse. The long-term goal of the work in this proposal is to identify, characterize and remove major sources of inefficiency in today's storage and compute infrastructures, and increase their utilization. The methodology will follow the PI's signature approach to research: form collaborations with industry to gain access to real data from production machines, analyze said data to characterize and identify problems, and use rigorous statistical and algorithmic methods to tackle core problems. The proposal outlines a set of specific near-term objectives that the PI identified in her recent work and plans to address over a 5-year horizon. She expects the results will open up new threads of research that will inform a longer term research agenda.