EAGER: Reconfigurable Network Hardware for Message-Driven Systems

EAGER：用于消息驱动系统的可重新配置网络硬件

• 批准号: 1446950
• 负责人: Douglas Swany
• 金额: $ 7.56万
• 依托单位: Indiana University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2014
• 资助国家: 美国
• 起止时间: 2014-10-01 至 2015-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1446950&HistoricalAwards=false
• 关键词: EAGER; Reconfigurable; Network; Hardware; Message

Message passing has long been known as a powerful tool for structuring parallel and distributed computation. Systems dating back more than 30 years have been built around message passing as the fundamental interprocess communications primitive. However, the use of message passing has been restricted to relatively coarse-grained application architectures due to message passing overhead. This project provides hardware to support an initial demonstration of coupling programmable network interfaces (10 Gbps NetFPGAs) into process scheduling in support of fine-grained message passing.The project explores novel hardware and software support for message-driven computation based on the concept of futures, a nearly 40 year old concept from programming languages. It also builds on both the NetFPGA platform and the memory models made available by the PCI-E bus and processors such as the latest Intel Xeon E5 series processors. Using the hardware support for locks and transactional memory as well as Data-Direct I/O into the cache in combination with a process execution model based on futures, this new approach to futures and message-driven scheduling revisits powerful historical approaches like dataflow in a realistic framework that stands to have impact in high-performance parallel computing and scalable network service architectures.

消息传递长期以来一直被认为是构建并行和分布式计算的强大工具。 30 多年前的系统都是围绕消息传递作为基本的进程间通信原语构建的。 然而，由于消息传递开销，消息传递的使用仅限于相对粗粒度的应用程序体系结构。 该项目提供硬件来支持将可编程网络接口（10 Gbps NetFPGA）耦合到流程调度中以支持细粒度消息传递的初步演示。该项目探索基于未来概念的消息驱动计算的新颖硬件和软件支持，一个来自编程语言的近 40 年历史的概念。它还构建在 NetFPGA 平台以及 PCI-E 总线和处理器（例如最新的 Intel Xeon E5 系列处理器）提供的内存模型之上。 使用对锁和事务内存的硬件支持以及进入缓存的数据直接 I/O 与基于 future 的流程执行模型相结合，这种新的 future 和消息驱动调度方法重新审视了强大的历史方法，例如数据流中的数据流将对高性能并行计算和可扩展网络服务架构产生影响的现实框架。