| With the development of cloud computing, Virtual Block Devices (VBDs) have become the mainstream solution to providing block storage for Virtual Machines (VMs) in large data centers. Compared with direct-attached disks, network-based VBDs have several remarkable advantages such as higher availability, scalability, and manageabil-ity. At the same time, with the emergence and development of Solid State Drives (SSDs), applications running within VMs that demand low I/O latency can benefit from these devices. However, due to the high unit capacity price, it is still exorbitantly expensive to build large data centers purely based on SSDs.In order to achieve both performance improvement and cost saving, hybrid storage systems which contain heterogeneous storage devices offer the ideal solution. Despite prevalent studies on hybrid storage systems, how to build cloud-scale systems that pro-vide hybrid block storage for VMs is still challenging concerning the following issues. Firstly, regarding virtualized workloads, I/Os that access unstructured Virtual Machine Disk Images (VMDIs) predominate, therefore making file systems not the best candi-date for storing such data. Secondly, the high availability required by VMs invalidates the offline data placement methods used in previous studies. Thirdly, the side effect caused by data migration is bound to be aggregated in cloud-scale data centers where unscalable data migration mechanisms are used.To address these issues, we designed and implemented a multi-tiered hybrid block storage system called MOBBS on the infrastructure of object-based storage. Leverag-ing the advantage of such storage architecture that dispersedly stores stripped data, MOBBS achieves an efficient use of SSDs by dynamically migrating data between dif-ferent storage tiers according to real-time workloads. An intelligent data placement method was designed to determine the appropriate object data layout. To relieve the side effect caused by data migration, this thesis proposes a parallel migration mecha-nism by which the burden of data migration can be distributed across the entire Object Storage Devices (OSDs), highly improving the scalability of our system. Moreover, since data are directly accessed from clients to OSDs in object-based storage, MOBBS is also highly suitable for storing unstructured VMDIs.Regarding system implementation, the source codes of Ceph and QEMU were modified so as to incorporate additional features. To avoid large changes towards exist-ing KVM-QEMU environments in data centers, we hid most complexities of MOBBS from QEMU while inserted them into Ceph. In the evaluation, both block level work-loads and file system workloads were used to assess the effectiveness and efficiency of our system. Furthermore, the effectiveness of the migration mechanism was evalu-ated. The results show that MOBBS manages to determine the appropriate data layout, perform non-disruptive placement reorganization, and thereby achieve an efficient uti-lization of hybrid storage devices. As the first study on hybrid block storage systems using object-based storage, this thesis will illuminate and inspire the researches on this field in the future... |
