Abstract: In the process of virtual clusters deployment, transmission of mass image data leads to longer boot time, and the scalability and efficiency are affected. An on-demand image distribution framework is proposed to reduce data transmission and avoid the bottlenecks of storage nodes and network based on the character that virtual machines only access limited data blocks in booting and running. Further, differentiated data block division method is proposed that takes into account the image accessing performance and distribution efficiency：the bigger division block size is used in deduplication processing to enhance index management efficiency and data accessing performance, whereas data in the boot region is divided with smaller data block size to improve data transmission efficiency. Distribution experiment of 50 image files shows that, compared with the full-image distribution, on-demand distribution can improve the distribution speed of 13~22 times.

Key words: virtual clusters, virtual machine, image distribution, on-demand distribution

摘要： 虚拟集群部署过程中，大量的镜像数据传输导致启动时间过长，可扩展性和效率受到影响。提出了镜像按需分发框架，根据虚拟机启动和运行过程中只对有限数据块访问的特性，仅传输虚拟机启动和运行所需数据块，避免了存储节点和网络瓶颈问题。进一步，提出了差异化的数据块划分方法，兼顾了镜像读写性能和分发效率：去冗余处理采用大数据块，提升管理效率和存取性能;采用小数据块划分启动区数据，提高传输效率。对于50个镜像的分发实验表明，相比于全镜像分发，按需分发速度可以提升13~22倍。

关键词: 虚拟集群, 虚拟机, 镜像分发, 按需分发