SHF: Medium: Bridging the Software/Hardware Gap Towards Efficient, Heterogeneous, and Predictable Datacenters

今天的大部分计算能力都存在于有时被称为“仓库规模计算机”的大型云计算基础设施中。这种转变催生了一类新的新兴应用程序，它们运行在数十万个强大的内核上，并访问PB或EB的存储空间;这些应用包括网络搜索、媒体流、大数据分析等。世界计算的这种集中化意味着这些系统中的低效率被放大到很高的程度--换句话说，如果我们能够提高这些系统的效率，我们就可以显著提高世界计算基础设施的效率（提高性能，减少能源消耗）。这项研究解决了效率低下的几个来源，包括越来越不准确的假设硬件同质性，不可预测的应用程序之间的干扰，以及低级别的资源共享模型差。本研究通过以下方式解决这些低效率问题：（1）为云平台创建一个异构感知执行框架，该框架不仅考虑到硬件的异构功能（预计会随着时间的推移而增加），而且有意采用异构性（在多个级别）提高运行不同工作量的效率;（2）创建用于共享资源管理的整体运行时系统，其考虑所有级别的资源共享，包括CMP和多线程核上的低级共享，允许线程被更积极地共同调度;以及（3）为性能和服务质量创建新的精确预测模型，其可以驱动更智能的调度决策。

Much of today's computational capability is housed in massive cloud computing infrastructures sometimes known as "Warehouse Scale Computers." This transition has given rise to a new class of emerging applications that run on hundreds of thousands of powerful cores, and access petabytes or exabytes of storage; these applications include web search, media streaming, big data analysis, etc. This centralization of the world's computing means that inefficiencies in those systems are magnified to a high degree -- in other words, if we can improve the efficiency of those systems, we measurably improve the efficiency (improve performance, reduce energy drain) of the world's computing infrastructure. This research addresses several sources of inefficiency, including increasingly inaccurate assumptions of hardware homogeneity, unpredictable interference between applications, and poor models of low-level resource sharing. This research addresses these inefficiencies by (1) creating a heterogeneity-aware execution framework for cloud platforms that not only accounts for the heterogeneous capabilities of the hardware that are expected to increase over time, but intentionally employs heterogeneity (at multiple levels) to improve efficiency in running diverse workloads; (2) creating a holistic runtime system for shared resource management that accounts for resource sharing at all levels, including low-level sharing on CMPs and multithreaded cores, allowing threads to be more aggressively co-scheduled; and (3) creating new precise prediction models for performance and quality of service that can drive more intelligent scheduling decisions.