Research On The Performance Isolation And Optimization In I/O Virtualization

With the advantages of consolidation, full resource usage, non-stop service, andperformance isolation, virtualization is widely used in data centers, while the expansionhas posed challenges of scalability from variable and unpredictable workloads. Usingvirtualization, logical resources could be simulated upon one single physical node and beallocated to various virtual hosts. In traditional data center, taking security and reliabilityinto consideration, each node possibly provides one service in one Operating Systeminstance. VMs provide suitable resources and portable environments to meet therequirements of modern servers. Although the I/O system in a virtual machine has a goodperformance, but with the increasing number of virtual machine in virtual platform, moreand more attentions have been paid to the research field of the scalability of virtualization.Performance of I/O system determines the performance of the entire virtual machineplatform. Unreasonable scheduling, management and distribution would impact thevirtual machine performance, affecting the users’ experience of services. Therefore, howto improve the scalability and performance of I/O system in virtual machine has becomean important issue.After analysis of the existing I/O virtualization system, this paper proposes severalworks in following areas: scheduling and management of I/O resources, optimization ofvirtualized block devices, and CPU scheduler for I/O requests latencies.The existing virtual machine management has some imperfections in I/O systemresources including network and block devices. For example, static allocation is notflexible, low utilization ratio of devices, and unable to meet the needs of performanceisolation for users. In order to solve these problems, the paper put forward a new resourcemanagement algorithm：Weighted Max-min fairness resource allocation. The algorithmdistributes each I/O resource based on the weight of virtual machine. Via the minimumguarantee user resources reserve, the algorithm ensures the normal running ofapplications in virtual machine. This algorithm has reached Pareto optimal by supportingproportional-share fairness subject to minimum reservations and maximum limitations on the I/O resource allocations for VM. Our results indicate that this weighted Max-minfairness resource allocation algorithm can effectively meet the difference needs of thevirtual machine application of I/O resources compared to original Credit resourcesallocation strategy. Dynamic adjust the weight on the I/O resource for VM improve theutilization of the I/O system hardware equipment, performance isolation and scalabilityof the I/O resources.Accessing to storage system in virtualization platforms suffers from too manyprotocol layers and data path, which leads to that access to storage devices relies highlyon the protocol processing speed and has a significant impact on resource overhead.Through using Xenoprofile on existing storage systems and analyzing proportion ofmodule in the critical path, a new driver architecture is proposed by parallel optimizingthe key module in the path to transfer the storage system bottlenecks. The experimentaldata shows that parallel optimization on the Xen block device storage drive can speed upthe system. Performance of the external storage device access is enhanced by20%compared with the existing Xen access architecture which means that the improvementeffectively improve the performance of I/O access in the virtual machine.CPU scheduling algorithm of the virtual machine can impact I/O system significantly.Based on EDF strategy, LEDF algorithm is proposed to improve the response time andthroughput of I/O request in virtual machine. LEDF algorithm improves the real-timeresponse of vCPUs task in the multicore architecture. The algorithm estimates thephysical CPU’s states by calculating the task of vCPUs. LEDF assigns vCPUs to PCPUsreasonablely according to the results and achieves load balance in the multicorearchitecture. Compared to Credit algorithm, LEDF algorithm saves50%request responsetime in the I/O intensive application environments and meets the needs of quality service.Based on the studies of I/O system in virtual machines, several innovative trendshave been put forward, which would be the foundation for the next researches.