Research On High Performance Overlay Network Embedded In Virtual Machine Monitor

As virtualization has become established and cloud computing emerges, many ofthe traditional very large-scale, scalable and highly reliable distributed computing havebeen moved to the purview of cloud computing and virtualization, and the resourceshave been shared and collaboratived in wide-area and dynamic environment, byaggregating the isolated and heterogeneous resources. For high-performance,high-density or mission-critical services and applications, using fault-tolerant ofvirtualization and cloud computing hardens against new vulnerabilities andcomplements the resilience.However, with the further expansion of the computing scale in cloud, the networkinfrastructure imposes signigicant and frequently unpredictable performance penalties.For the tightly-coupled parallel applications, especially, the system efficiency would beloss if the network performance became lower. The virtual computing environmentcomposed by virtual machines (VMs) and virtual overlay network in distributedcomputing and cloud computing, can achieve the advantages of dynamic networktopology, performance prediction and virtual environment migration. For highperformance distributed computing environment, a new simple and flexible model hasbeen designed and implemented based on virtualization technology, in whichtightly-coupled applications can seamlessly migrate to and from differentnetworks. This work has important significance in cloud computing system or datacenter. The main innovations of this paper include:1. An overlay network based on the virtual machine, has been researched anddesigned. The overlay system, VNET/P, is embedded into Palacios virtual machinemonitor (VMM). Virtual overlay that implemented in this paper first integrated withvirtual machine monitor at the research field of virtualization and overlay. VNET/Pabstracts a LAN for a collection of VMs regardless of their current locations, and itencapusulates, routes and forwards data based on the MAC in layer2and routing rulesthat user setted. The test proved that the system has negligible latency and bandwidth overheads in1–10Gbps networks.2. A unique bridging system has been designed, linking the local VMs withoutside network. The VNET/P transmits data from or to external network based on thebridging system. In Palacios+Kitten context, the virtual bridging device is located in abridge VM Dom0, and conveys packets using the Virtio virtual NIC. The bridgingsystem re-encapsulates and sends the packets to the Dom0on destination host throughthe Passthrough Ethernet device.3. The data have been transmited parallelly between VMs, applying multi-coretechnology to the virtual network. In multi-core mode, if VNET/P needs to receive orsend a package, the dispatcher made a kernel thread and then polled a virtual card. TheVM supported by VNET/P assigns a CPU core. Here, the CPU does not execute theVM exit, but VMs run parallelly on some different CPU core. Using this approach todynamically employ idle processor cores, to reduce the VM exit and to increase packetforwarding bandwidth.4. A virtual overlay network has been established on VNET/P node. The VNET/Pis enhanced to become an adaptive virtual network that can dynamically modify itstopology, monitor network performance and migrate VMs. The toplolgy is not just asimple Ethernet topology, but can be dynamically configured general topology, such asbus, star, ring, mesh and cube structure. VNET/P based on the MAC address candirectly interact with some physical network devices, virtual NICs in the VMM andfast virtual devices in Guest. In addition, this paper formalizes the adaptation problem.