【作者】 罗伟；

【导师】 余辰；

【作者基本信息】 华中科技大学 ， 计算机技术， 2016， 硕士

【摘要】 云计算的IaaS(Infrastructure as a service,基础设施即服务)模式采用了向用户提供虚拟机的服务形式,虚拟机在云计算下扮演着至关重要的角色。英特尔公司实现的硬件级虚拟化技术,以及Xen或KVM等虚拟机管理开源软件,很大程度上降低了虚拟机自身的实现开销。随着云数据中心规模的不断扩大,虚拟机镜像存储系统成为了非常重要的一个部分,它关系着虚拟机的部署开销以及应用的运行时性能。不同于传统数据中心采用的SAN等集中式存储方案,基于多级缓存的网络自适应虚拟机镜像存储系统MADStor(Multi-level cache-based and Adaptive Distributed Storage)采用了分布式协作缓存的架构,除了降低成本外,也不会受单点网络带宽等因素影响。另外,MADStor也针对目前主流的协作式存储系统,优化了云数据中心网络负载不均衡的现象,避免因此而产生的计算节点I/O瓶颈。MADStor在计算节点上采用了两级磁盘缓存的策略,很好地解决了数据共享以及脏数据现象避免的问题。同时,为降低计算节点存储资源消耗,系统采用了基于不同策略的双重去重技术。MADStor优化了分布式对象存储系统中的寻址技术,针对节点网络负载情况动态调整基于对象名的映射信息,同时将基于对象内容的映射信息存于内存缓存中,进一步提升了虚拟机的I/O性能。MADStor系统的实验结果表明,与传统的集中存储方案以及主流分布式协作方案相比,均获得了良好的性能提升,尤其在网络负载不均衡的情况下,虚拟机的部署时间可缩短2-3倍,运行I/O密集型应用时的吞吐率可提高4-7倍。更多还原

【Abstract】 IaaS(Infrastructure as a service) models in clouds provide services for users in the form of virtual machines. Virtual machines play a significant role in cloud computing. Hardware virtualization technology implemented by Intel and virtual machine management, such as Xen or KVM, largely reduced the implementation cost of the virtual machine itself. With the continuous expansion of cloud data center, virtual machine images storage system has become a very important part. It has a great influence in both deploying cost of virtual machines and run-time performance of applications.Different from the solutions adopted in traditional data center, where SAN or other centralized storage are deployed, MADStor(multi-level cache-based virtual machine image storage system) uses a distributed cooperative caching architecture. In addition to reducing costs, it can avoid being affected by network bandwidth bottleneck in single node. Furthermore, MADStor optimizes network load imbalance in compute nodes, comparing the general collaborative storage system, to avoid I/O bottlenecks. MADStor adopts two disk caches on compute nodes, which brings data sharing and avoids the phenomenon of dirty data. Meanwhile, double deduplication strategies are proposed to reduce storage consumption in compute nodes. MADStor optimizes chunk addressing techniques in distributed object storage system. According to network load distribution, mapping information based on the object name would be dynamically adjusted. And, MADStor stores the mapping information in memory based on contents of objects to increase virtual machine I/O performance.Experimental results show that MADStor has a better performance, compared with traditional centralized storage and the mainstream distributed collaborative system, especially in the situation that the network load is distributed unevenly.更多还原