Research On Optimization Of Shared Memory Mechanism Towards High Communication Performance Between Virtual Machines

On cloud platform such as a large-scale data center,network I/O is a significant part of system workload.Performance improvement on network I/O between Virtual Machine(VM)is a research hot spot wildly concerned by both academic and industrial community.One effective and popular network acceleration method is using shared memory channel to bybass network between VMs located on one physical machine,while still using TCP/IP protocol stacks to communicate between VMs located on different physical machine.The research work of this thesis bases on a transparent co-resident aware inter-VM high performance communication mechinization XenVMC,which is designed and implemented by our research group.The main issues include improvment of co-resident membership maintanance and enabling multi-core to realize concurrent reading and writing on shared memory.The main achievement of this thesis includes:1.introduce the latest technology development situation of cloud platform,study the main research hot spots on cloud network architecture nowadays,analysis the latest research trend in network I/O virtulization.With a basis introduction on network I/O virtulization technologies,this thesis studies the existed related works on software based network I/O virtulization in detail,and concludes these works' properties and deficiency,accounting for the background and significance of our research work.2.It is a wildly used solution to accelerate communication between VMs located on same physical machine by enabling shared memory channels.The open-sources and influential related works include XenSocket,XWay and XenLoop.While XenSocket dose not support transparence to user-level applications.The applications have to be re-writed before running.On the contrary,XWay and XenLoop support application transparence.This thesis transplants XWay and XenLoop to the same running environment with XenVMC.By comparing the communication performance with XWay and XenVMC using a typical network performance benchmark netperf,XenVMC has lots of advantages both on design and implementation.3.Co-resident membership maintanance is a basic issue to use shared memory to accelerate communication between co-resident VMs.The existed polling based methods collect co-resident membership and update membership at a fixed cycle.No matter co-resident membership changes or not,every cycle the polling based method spends CPU and network resource on membership maintanance.Besides,themembership updation of polling methods happens at the end of every cycle,which means the membership changes can not be proceed at once.Based on these disadvantages of polling based methods,this thesis proposes an event driven co-resident membership maintanance method,called CoKeeper.CoKeeper updates membership at the right moment when co-resident relationship changes by an incremental approach.In design,CoKeeper reaches four goals: low overhead,transparency,live migration support and high response speed.Comparing with one polling based dynamic method,experiment results show that CoKeeper has advantages on overhead,VM amount scalability and response speed.4.Multi-core technology is wildly deployed on cloud environment,which implies the possibility of improving communication performance by enabling multi-core optimization.Most of related research works implement data read/write between VMs with a lock-free buffer,while under multi-core situation,these buffers do not support concurrent data reading or concurrent data writing.This thesis proposes multi-core optimization to XenVMC.By changing data receiving process,the multi-core optimization allows multiple readers receive data from shared memory channels concurrently.Experiment results show that after optimization,UDP throughout of XenVMC increases to about 1.2 times,transcations of UDP and TCP increase to about 1.5 times.