CSR--PDOS: Exploring Cluster-Based Data Center Design Space for High Performance and Dependability

集群已经成为设计高性能和可靠的数据中心的最具成本效益的解决方案，这些数据中心越来越多地被部署用于各种基于Web的服务。然而，数据中心对整体延迟有很大的贡献，随着动态Web内容的使用越来越多，这可能会增加。此外，托管关键服务的（24 x7）服务器可用性似乎是一个遥不可及的目标。因此，设计高性能、高可靠的数据中心成为一个重要的问题，本研究的目的是研究一个由前端Web服务器、中间层应用服务器和后端数据库服务器组成的三层数据中心的设计，InfiniBand架构（IBA）通信范例。原型数据中心将在96节点IBA上实现，连接的Linux集群，使用真实的应用程序和基准测试来研究三个研究问题。这些问题是量化使用用户级通信技术（如IBA）代替标准TCP/IP协议栈的性能优势，提出新的数据中心性能增强技术，以及表征不同故障对服务器可靠性的影响。该研究将有效地利用IBA通信机制来设计多层数据中心，量化使用IBA的端到端性能优势，为高可用性数据中心开发容错机制，并了解高性能和可靠性之间的设计权衡。除了在数据中心设计的几个领域培育新的研究方向外，这项研究的成果很可能具有重大的商业利益。

Clusters have emerged as the most cost-effective solution todesign high performance and dependable data centers,which are increasingly being deployed for a wide variety of Web-based services.However, data centers contribute to a significant part of theoverall delay, which is likely to grow with the increasing useof dynamic Web contents. Furthermore, (24x7) server availabilityto host critical services appears to be a far-fetched objective. Therefore, design of high performance and dependable data centershas become a critical issue.The objective of this research is to investigate the design of a three-tierdata center consisting of a front-end Web server, a middle-tier applicationserver, and a back-end database server using the industry standard, InfiniBand Architecture (IBA) communication paradigm.A prototype data center will be implemented on a 96-nodeIBA-connected Linux cluster to investigate three research issuesusing real applications and benchmarks. The issues are quantifying the performance benefits ofusing a user-level communication technique such as IBA instead ofthe standard TCP/IP stack, proposing new performance enhancementtechniques for data centers, and characterizing the impact of differentfaults on server dependability.The research will enable effectively utilizing the IBA communication mechanismfor designing multi-tier data centers, quantifying the end-to-end performancebenefits of using IBA, developing fault-tolerant mechanisms for highly availabledata centers, and understanding the design tradeoffs between high performanceand dependability.In addition to fostering new research directions inseveral areas of data center design, the outcome of this researchis likely to have significant commercial interests.