The Research And Implementation Of Metadata-Manage Staregy Technology Based On The Object Storage System

The demand of high I/O throughput，high parallel data transfers and high expansibility insuper computing is increasing imminently which drive the storage systems must make changesbecause the traditional storage systems cannot adapt to the performance requirements. In order toachieve good storage characteristic, Object-based Storage System (OBS) have been introducedto fit the application demand which have been the development direction in research of storagetechnology.In OBS，file-level metadata is mainly managed and stored by the Meta-data Server (MDS),and the metadata has the property of small size and mass amount. The capability of the storagesystem in the OBS operation system is determined by the management efficiency of the metadatabecause50%to80%file operations correlate to the metadata. There are some problems inmetadata management strategy:1)The rename of the metadata and directory modification lead tomass metadata migration and network spending.2) Centralized single MDS has become thehotspot which cause to load unbalance and the performance decrease of the system because ofthe highly access to the same metadata by mass users at the same time.To solve the problems above, this dissertation proposed a new metadata managementstrategy called DMHM which can be deployed in distributed multi-MDS environment. DMHMintroduced an improved B+tree structure which divided the directory subtree into B+subtrees tomake the management and storage of the directory metadata and file metadata separately.DMHM not only reduced the spending of the rename and modification but also avoided theproduce of the hotspot to keep the load balance. Based the method above, designed a smalldistributed MDS model called DMSS, and implemented the DMHM operations of the file,directory, list and buffer in the DMSS.This dissertation designed a dynamic balance algorithm which introduced the hotspotstatistic concept to solve the load balance problem. The hotspot statistic are focused on the leafmetadatea when use the DMHM. The algorithm can make the operation of add, separate, copyand delete with the hotspot collected by the two strategies, and enhanced the availability andexpansibility of the system.Finally, static sub-tree method, Lazy Hybrid (LH), France and DMHM strategy are testedand analyzed. Test of distribution of metadata showed that metadata of DMHM strategy ingeneral distributed evenly; test of metadata mobility show that the DMHM strategies migrationrate is lower; test of throughput show that DMHM strategy has relatively obvious advantages toother the two strategies, more suitable for dynamic changes in distributed MDS system; test of operation time respond for reading, creating operation and renaming show that the systemresponse time of the DMHM strategy was significantly lower than the static sub-tree method,and improvement is also very clear in the rename operation of file directory, has a greatadvantage.