CPA-CSA: Virtualization-Aware Architectures to Accelerate Network I/O Processing

CPA-CSA：加速网络 I/O 处理的虚拟化感知架构

• 批准号: 0811834
• 负责人: Laxmi Bhuyan
• 金额: $ 30万
• 依托单位: University of California-Riverside
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-08-01 至 2013-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0811834&HistoricalAwards=false
• 关键词: CPA; CSA; Virtualization; Aware; Architectures

Proposal ID: 0811834 Title: Virtualization-Aware Architectures to Accelerate Network I/O Processing PI name: Bhuyan, Laxmi N. Institution: University of California-Riverside ABSTRACTThe current application servers consist of high-performance computing, web and e-commerce servers, proxies, database servers, and multimedia servers, all of which need high rates of network I/O processing. The protocol-processing overhead in these servers dominates the CPU execution time. Power and thermal constraints have forced CPU vendors to turn to core multi-processing (CMP) instead of frequency scaling to increase performance. Complimentary with CMP is surging industry interest in virtualization for diverse uses including server consolidation and sharing of various resources. However, several studies have shown that virtualization brings significant extra overhead to network I/O. The project explores new architectural techniques in virtualized environment that reduces the network I/O overhead through the following approaches:? Measurement techniques are developed to trace the life of a packet in a virtualized environment with multi-core server and 10 Gigabit Ethernet. Architectural statistics like instruction types, IPC, and cache misses, etc. are measured and compared with the results of the guest O/S protocol. The study reveals architectural bottleneck points in network I/O processing.? Techniques are developed for intelligent interrupt scheduling, TCP onloading, affinity-based core scheduling, and domain-to-core-mapping to show that the performance is enhanced to a great extent.? The project implements several architectural techniques, such as ISA optimization for protocol stack, incorporation of network cache and direct cache access, memory copy engines, and CPU+NIC integration; and proves their superiority in performance of a virtualized I/O system. ? In order to evaluate new architectural ideas, a cycle-accurate full system simulator is designed that can evaluate virtualized IO operations. The simulation software is made available to researchers.

题目：加速网络I/O处理的虚拟化感知架构PI名称：Bhuyan， Laxmi N.机构：加州大学河西分校摘要当前的应用服务器包括高性能计算服务器、web和电子商务服务器、代理服务器、数据库服务器和多媒体服务器，这些服务器都需要高速率的网络I/O处理。这些服务器中的协议处理开销占CPU执行时间的大部分。功率和热限制迫使CPU供应商转向核心多处理（CMP），而不是通过频率缩放来提高性能。与CMP互补的是，业界对各种用途（包括服务器整合和各种资源共享）的虚拟化的兴趣日益高涨。然而，一些研究表明，虚拟化给网络I/O带来了显著的额外开销。该项目探索虚拟化环境中的新体系结构技术，通过以下方法减少网络I/O开销：？在多核服务器和10千兆以太网的虚拟化环境中，开发了测量技术来跟踪数据包的生命周期。诸如指令类型、IPC和缓存丢失等体系结构统计数据被测量，并与客户机O/S协议的结果进行比较。该研究揭示了网络I/O处理中的架构瓶颈点。开发了智能中断调度、TCP加载、基于亲和力的核心调度和域到核心映射等技术，表明性能得到了很大程度的提高。该项目实现了几种架构技术，如协议栈的ISA优化、网络缓存和直接缓存访问的合并、内存复制引擎和CPU+NIC集成；并证明了它们在虚拟I/O系统性能上的优越性。？ 为了评估新的架构思想，设计了一个周期精确的全系统模拟器，可以评估虚拟IO操作。仿真软件已提供给研究人员使用。