EAGER: Preliminary Study to Demonstrate Feasibility and Advantages of Massively Parallel Server Processors

今天的数据中心一直面临着实现更高性能和提高能源效率的压力。然而，在现代数据中心中消耗大部分能源的服务器处理器，在十多年的时间里，在性能和效率方面只取得了渐进式的改进。这个建议通过研究一种新的处理器架构来解决问题的核心。从头开始设计，考虑到现代服务器工作负载的特点&amp；#8213；这可以为未来的数据中心提供前所未有的性能和能源效率。这项工作利用了服务器经常并发处理类似请求的观察结果，并在请求之间执行相同的指令序列。这意味着可以利用单指令多线程（single - instruction multiple - threads， SIMT）处理器设计来充分利用服务器的请求级并行性，从而大幅提高服务器处理器的效率。不幸的是，现有的SIMT架构&amp；#8213；最值得注意的是图形处理单元（gpu）&#8213；由于服务器工作负载的严格服务质量要求，从根本上不具备利用这种请求级并行性的能力。因此，基于simt的服务器处理器需要明显不同的设计。该项目研究了将SIMT范式应用于计算堆栈不同层的服务器处理器的可行性和挑战，从微架构到内存层次再到系统软件。该项目探讨了SIMT服务器处理器上不同硬件和软件设计技术的性能影响，以确定对此类处理器最有利的特性。研究结果将为显著提高服务器处理器的能源效率开辟一条道路，并为计算机体系结构和系统社区开辟新的研究方向。

Today's data centers are under constant pressure to achieve higher performance and improved energy efficiency. However, server processors, which consume the lion's share of energy in a modern data center, have only made incremental improvements in performance and efficiency in more than a decade. This proposal gets at the heart of the problem through investigating a novel processor architecture―designed from scratch with characteristics of modern server workloads in mind―that can provide unprecedented levels of performance and energy-efficiency to the data centers of the future.This work leverages the observation that servers frequently service similar requests concurrently, and execute the same instruction sequences across requests. This implies that the abundant request-level parallelism of servers can be harnessed with a Single-Instruction-Multiple-Threads (SIMT) processor design to achieve drastic improvement in server processor efficiency. Unfortunately, existing SIMT architectures―most notably, Graphical Processing Units (GPUs)―are fundamentally ill-equipped to exploit this request-level parallelism due to stringent quality of service requirements of server workloads. Therefore, significantly different designs are required for SIMT-based server processors. This project investigates the feasibility and challenges of applying the SIMT paradigm to server processors at different layers of the computing stack, from micro-architecture to memory hierarchy to system software. The project explores the performance effect of different hardware and software design techniques on SIMT server processors to identify the features that benefit such processors the most. The results will lay out a path toward significantly improving the energy efficiency of server processors, and open new research directions for the computer architecture and systems community.