CAREER: Leveraging temporal streams for micro-architectural innovation in data center servers

随着在线服务的全球用户基础不断扩大，新服务和新功能迅速开发，这些在线云服务运营所在的数据中心面临着实现更高性能和改善服务质量的持续压力。然而，支持云服务在人们生活的各个方面的采用？S日常生活需要将数据中心扩展到极端规模，拥有数亿台服务器和生态上不可想象的能源账单。这项研究开发技术来提高未来数据中心的性能和效率，目标是从每台部署的服务器获得更高的性能和更低的能源成本。因此，它直接促进数据中心和在线服务的可持续增长，同时培养专门处理未来数据中心和云所面临的挑战的世界级专家。这项研究利用最近编码的一种被称为“时间流”的现象来解决云中服务器系统面临的一些长期存在的微体系结构性能瓶颈。在过去几十年中为台式机、移动和超级计算机域开发的许多性能增强技术对服务器系统的好处有限，因为典型云工作负载的大小和复杂性需要比这些技术目前可用的元数据存储容量大得多的元数据存储容量。这项工作重新构建了投机性结构的元数据存储，利用时态流来扩展它们的有效容量。具体地说，这项工作以指令预取器、分支预测器和硬件存储器为目标作为案例研究，以展示时态流在执行云工作负载时为这些机制提供足够的元数据存储的能力。

As the global user base for online services continues to expand and new services and features are rapidly developed, data centers from which these online cloud services operate experience constant pressure to achieve higher performance and improve their quality of service. However, supporting the adoption of cloud services in all aspects of people?s daily lives requires expanding data centers to an extreme scale, with hundreds of millions of servers and ecologically unthinkable energy bills.  This research develops technologies to improve the performance and efficiency of future data centers, targeting higher performance and lower energy costs from each deployed server.  As such, it directly contributes to sustainable growth of data centers and online services, while at the same time training world-class experts specialized in tackling the challenges facing future data centers and clouds.This research leverages a recently-codified phenomenon called "temporal streams" to solve a number of long-standing micro-architectural performance bottlenecks facing server systems in the cloud.  Many of the performance enhancing techniques developed over the course of the past several decades for the desktop, mobile, and super-computer domains provide limited benefits to server systems, because the size and complexity of a typical cloud workload requires significantly greater meta-data storage capacity than currently available to these techniques.  This work re-architects the meta-data storage of speculative structures, leveraging temporal streams to expand their effective capacity.  Specifically, this work targets instruction prefetchers, branch predictors, and hardware memorization as case studies to demonstrate the ability of temporal streaming to provide sufficient meta-data storage for these mechanisms when executing cloud workloads.