Optimization Of Image Storage And Transmission For Virtual Clusters

The adoption of virtual cluster technologyimprovesresource utilization rate, at the same time,performance ofthe storage systemis also madehigh demands.On the one hand, the introduction of the sparse image filewhich is amain storage formofvirtualmachines, increased fragmentation ofthe storage system; on the other hand, multiple virtual machinesrun simultaneously ona single physical machine, and thencompetition of differentaccessmode makes I/Operformance of the storage system to decline further.In addition, the progress of virtualclustersdeployment and migration involvesstorage anddistribution fora large number ofimage file, it is not easy for efficient storageandtransmission ofthese imageswhen it is needed.Based on the researchand analysis of the I/Oworkflow characteristics in multiple virtualmachines environment, thethesis aims at aboveproblem, and carries out research. The main works and contributions include:We designe a hybrid storage architecture for virtual machine image, and propose the hot zone tracking replace algorithm according to the characteristics of multiple virtual machine workflow. The hybrid storage architect ure consists of a solid-state driver(SSD) and a general hard driver disk(HDD), by taking advantage of the outstanding random access performance of SSD, overcome the I/O performance degradation problem caused by increased randomness of workload in multiple virtual machines environment.Thehot-zone tracking replacement algorithm uses an efficient data structure to maintain th e hotness of global data block, and ensures that the data of hot zone reside in the SSD.Experiments show that, in multiple virtual machines environment, compared with the traditional algorithm the hybr id storage system with hot zone tracking replacement algorithm can improvethe hit rateof 12%, using the appropriateprefetch strategycanimprovethe hit rate of 23%, while the cumulative I/Oresponse timedecreased by 34 %.We propose a framework of virtual machine image on-demand distribution. According to the accessed data blocks are usually limitedduring the booting up and running of virtual machines, On-demand distribution mechanism reduced the data transfer in the cluster deployment and closing process, to avoid bottlenecks of image storage node and network brought by the large number of access requests. We disigned differentiated data block size division method, which divides image data for deduplicationusing bigger data block size, and divides image data by smaller data block size for transmission which can improve the efficiency of data. Read and write performance of deduplication image storage system and image data distribution efficie ncy is taken into account. Distribution experiment of 50 image files shows, compared with the full-image distribution, on-demand distribution can improve the distribution speed of 13-22 times.We propose virtual machine fast live migration architecture for multi-dataceter and corresponding migration method, the main idea is: to reduce the data transfor of virtual machine migration by use of copies of images, this make virtual machine can be rapidly migrated between datacenters through low-bandwith WAN; data transfer can be reduced further using of fast live migration cross datacenter algorithm which combinated pre-copy and post-copy method and dirty data block deduplication; remote data block fault rate can be reduced by mechanism prefetching accordingof data block fault and hot-zone ordered transmission, and effect on virtual machine performance during image migration can also be deduced.Experiments show that, compared to the previoustraditional algorithm migration time ofoptimized Flim CD algorithmis decreasedby 64.3%,and the lack block rate in image postcopy phasedropped more than49%.