云数据中心 I/O 资源池化

近年来,云计算的发展为数据中心带来了新的应用场景和需求.其中,虚拟化作为云服务的重要使能技术,对数据中心服务器I/O系统的性能、扩展性和设备种类多样性提出了更高的要求,沿用传统设备与服务器紧耦合的I/O架构将会导致资源冗余,数据中心服务器密度降低,布线复杂度增加等诸多问题.因此,文章围绕I/O资源池化架构的实现机制和方法展开研究,目标是解除设备与服务器之间的绑定关系,实现接入服务器对I/O资源的按需弹性化使用,从根本上解决云计算数据中心的I/O系统问题.同时,还提出了一种基于单根I/O虚拟化协议实现多根I/O资源池化的架构,该架构通过硬件的外设部件高速互连接口多根域间地址和标识符映射机制,实现了多个物理服务器对同一I/O设备的共享复用;通过虚拟I/O设备热插拔技术和多根共享管理机制,实现了虚拟I/O资源在服务器间的实时动态分配;采用现场可编程门阵列(Field-Programmable Gate Array)构建了该架构的原型系统.结果表明,该架构能够为各个共享服务器提供良好的I/O操作性能.

I/O Resource Pooling for Cloud Datacenter

The emergence of cloud computing brought datacenter to new application environments and new requirements. Server virtualization, as the main enabling technology for cloud services, needed datacenter to increase the number, performance and variety of I/O devices to support more virtual machines, for which the traditional server I/O architecture could result in waste of resources, increasing of server density and rise of wiring complexity. I/O resource pooling was a promising architecture to solve these problems, by separating devices and servers and enabling elastic on-demand provisioning of I/O resources to servers. We designed and implemented a single root I/O virtualization based multi-root I/O resource pooling architecture. By using a hardware-based peripheral component interface express identifier re-mapping, different physical servers can share the same physical devices, and through a hot-plug based allocation method, virtual I/O device could be dynamically reallocated between servers. The correctness and performance of our pooling architecture was proved by a Field-Programmable Gate Array prototype in this paper.