Network Performance Optimization Of Virtual Machine

With the rapid development of virtualized technology and the improvement of capacity of CPU, the more guest virtual machines (VMs) the virtualization hosted platform supports, the more services and applications the platform provides, and users make new requirements on the performance of network and resource allocation of network. In virtualized environment, to meet the high requirement of network performance and user demands on resource allocation, it is necessary to investigate methods on improving network performance and allocation network resources. This thesis uses a multi-level and multi-granularity performance combinative evaluation method to evaluate and analyze network performance of virtualization. Then, based on the evaluation and analysis on network performance of virtualization, the thesis studies the network performance of virtualization and management of network resource from many aspects. These aspects include the performance optimization of virtualized network based on a model of network performance, fine-grained performance isolation of virtualized network, and distributed management of virtualized network resource for quality of service (QoS) of application.The evaluation and analysis of virtualized network performance is the fundamental prerequisite and key step, for the performance optimization of virtualized network. According to the issues on the evaluation of virtualization, the thesis proposes a multi-level and multi-granularity performance combinative evaluation method to identify effectively the performance bottlenecks of virtualization systems. In this approach, benchmark tests cooperating with trace method gets different grained results in different levels, to describe the performance of virtualization perfectly. According to this approach, designing and implementing a test tool, named VSCB, is to configure flexibility testing environments and deploy automatically testing tasks. Besides, to cover the shortage of virtualization trace method, VSCB presents Xenrelay, a unified, efficient, and simple mechanism for transferring large amounts of data from guest domain kernel to privileged domain user-space, for optimizing performance on transferring trace data. Furthermore, the evaluation of performance of virtualized network with this approach shows that network performance of virtualization is need to improve and the current state of performance isolation for virtualization is however much more rudimentary.To improve the network performance of virtualization, a network model of virtualization is presented. This model, with queuing network theory, presents same theoretical methods to optimize the network performance of and its practical significance. In the virtualized environment, to meet the requirement of network grows with the increase of the number of VMs, the network model shows that parallel transmission of data with multi-NICs improves network performance. However, VMs hardly make best of multi-NIC, leading to low efficient data transmission under multi-NIC and a waste of network resources. To solve this issue, the thesis, based on the analysis of network model, proposes a mutli-NIC performance optimization strategy of virtualized network. This strategy provides a network parallel module by expanding the ARP neighbor system, to achieve the expansion map IP address of VM to MAC addresses of many physic NICs. Hence, the VM transmits data parallel via multi-NICs by network parallel module. At the same time, the strategy also introduces transmit and receive load balancing mechanism between Multi-NICs, for making full use of every NIC.To ensure performance isolation among VMs, this thesis presents a fine-grained resources allocation framework applied in general I/O device of virtualization. The framework, based on abstracted Xen I/O model, provides both throughput and fairness guarantees for network and storage device via2-level resource control, for ensuring the fain-grained I/O performance isolation among VMs. The high-level tier makes guest domains perceive the states of the resource through resource control on shared ring buffer. Besides, the high-level tier uses public/private token-bucket to support work-conserving, and try to fairly distribute any spare bandwidth to the different VMs with a compensated allocation. The low-level tier is intended to meet the fairness guarantee, and computes token quantity value of each bucket by feedback-driven scheduler. Last, to prove the effectiveness of the framework, implementing the protosystem of resources allocation framework is to allocate network resource in the virtualized environment, and the results of experiments shows that the framework ensures the network performance isolation among VMs.Virtualized technology improves the utilization of resources and saves energy, so that more and more users of high performance computer use virtualized technology to build distributed environments. In the distributed environment with virtualization, the thesis describes a distributed resource allocation of VMs’ network for application QoS. This resource allocation is a system based on a forecasting model of network resource allocation. In this resource allocation, sending end forecasts the resource allocation of receiving end, according to historical data of network resource allocation which receiving end transferred. And then, based on the forecasting information and network usage of sending end, the sending end control the transmission of data effectively, which guarantees network QoS of VMs in the distributed environment.In brief, based on the evaluation and analysis of virtualization with a multi-level and multi-granularity performance combinative evaluation method, the thesis outstandingly researches a serial of mechanisms and implementation methods of virtualized network to improve virtualization network performance, guarantee performance isolation between VMs and rational allocate network resources in virtualization environment. Especially, this thesis mainly researches some key technologies including network performance model, network parallel transmit mechanism with Multi-NICs, fine-grained performance isolate for network, and distributed network resources allocation and so on, to offer a series of technologies and solutions for the issues of network subsystem in virtualized environment.