Performance Analysis of a Fibre Channel Switch supporting Node Port Identifier Virtualization
The server virtualization architecture encompassing sharing of storage subsystems among virtual machines using fibre channel fabrics, to improve server utilization and reduce the total cost of ownership, was pioneered by IBM through their System z9 mainframe and its predecessors. With the advent of sharing small computer system interface storage subsystems among host servers through fibre channel based storage area networks, has cropped up new set of security and associated performance issues when the host servers are virtual machines on a single physical server. To address the security issues and reduce the total cost of ownership, IBM introduced new storage virtualization architecture known as node port identifier virtualization enabling thousands of virtual machines on a server to share storage subsystems through a few numbers of host bus adapters.
In this paper, we introduce the node port identifier virtualization architecture and the associated fibre channel switch latency performance issue that would affect virtual machine instantiation when supporting thousands of virtual machines. We first show the architectural problem in hard zoning mechanism contributing to the large fibre channel switch latency by actual performance measurements on a switch using hardware simulators. Next, we suggest a modification to the hard zoning mechanism to reduce the fabric channel switch latency significantly and demonstrate the reduction using hardware simulators. The performance issue we have identified and addressed will allow a single fibre channel switch to support thousands of virtual machines on a server using only a few numbers of host bus adapters.
- There are currently no refbacks.