Network Interface Support for High-Performance and Scalable Communication Services in Clusters

集群计算系统正变得越来越流行，因为它为各种应用程序（如数据库、数据挖掘、可视化、协作计算和高性能计算）提供经济有效且负担得起的计算环境。这些应用程序需要来自底层通信子系统的四种基本通信服务：点对点通信、集体通信、同步和服务质量（QoS）。用于集群的现代网络接口卡（nic）包含一个或多个可编程处理器、DMA通道和大量内存。网卡变得更快、更智能、更智能，并被指定为活动网络接口。建议沿着以下方向进行研究，以利用这些活动网络接口为下一代集群实现和提供可扩展的通信服务：1)利用多cpu网卡及其架构特性提供快速点对点通信，2)设计网卡级支持高性能和可扩展的集体通信和同步，3)利用可编程网卡级特性提供QoS支持，4)使用消息传递编程模型实验评估新/增强的通信服务对广泛应用的好处。通过广泛的理论、实验和仿真研究，计划为具有Myrinet和千兆以太网互连的集群设计、开发、实施和评估基于nic的解决方案。

Cluster computing systems are becoming increasingly popular forproviding cost-effective and affordable computing environments for arange of applications (such as databases, data mining, visualization,collaborative computing, and high performance computing). Theseapplications require four fundamental communication services from theunderlying communication subsystem: point-to-point communication,collective communication, synchronization, and Quality of Service(QoS). Modern Network Interface Cards (NICs) for clusters incorporateone or more programmable processors, DMA channels, and large amount ofmemory. The NICs are becoming faster, smarter, and intelligent and arebeing designated as Active Network Interfaces. Research is proposedalong the following directions to take advantage of such activenetwork interfaces for implementing and delivering scalablecommunication services for next generation clusters: 1) Usingmulti-CPU NICs and their architectural features to provide fastpoint-to-point communication, 2) Designing NIC-level support forhigh-performance and scalable collective communication andsynchronization, 3) Taking advantage of programmable NIC-levelfeatures to provide QoS support, and 4) Experimentally evaluating thebenefits of the new/enhanced communication services for a wide rangeof applications using message passing programming model. Throughextensive theoretical, experimental, and simulation studies, NIC-basedsolutions are planned to be designed, developed, implemented, andevaluated for clusters with Myrinet and Gigabit Ethernetinterconnects.