Researchs Of Replication Management On Object Storage System

Object-based Storage technology can fully make use of its Object Storage Device (OSD)'s function and processing capacity, and transfer part of systemic function to those devices. It can provide object APIs which are rich in expression and thus establish the system's extendibility, share quality, high availability and reliability.In the internet environment, comparatively it covers more OBS clients, more client types, more complicated network circumstance and more storage object accordingly. This actually demands a higher level of service in terms of reliability and high performance to be fulfilled by the OBS. To conquer the problem, this paper proposes a research on replica-based Object Storage System. Through replica created with placement in proper storage node, the replica can response to partial I/O requests towards the original object and thus realizes the nearest answer principal. Eventually, it provides reduced access latency, bandwidth saving and better overall performance.Nevertheless, the introduction of replica technology to the system would result in redundant versions of object replicas within OBS. These replicas are so significant to effective management that inappropriate monitor strategy can lavishly consume the storage capacity and result in unexpected downhill performance. To avoid this phenomenon and sufficiently utilize the replicas, this paper has studied the key technologies in replica-based OBS system, which can be listed as follows:(1) We propose the possible implementation of replica technology in Object-based Storage system, and gives three various replica-based OBS frameworks. They are the centralized, decentralized and collaborative frameworks. Centralized management framework means that the metadata server takes the full-load job to manage those replicas in the system. This mode is pretty easy to realize but unavoidably, the metadata server itself can easily be a bottleneck. On the other hand, decentralized framework refers to a mode that replica management is exclusively implemented by every OSD, so the chance of bottleneck is largely tone down. Unfortunately, the frequent communications between OSD in this framework has resulted in marvelous system consumption. Based on these previous two, we propose a new collaborative architecture which places the systemic information control like replica locating or balancing on the metadata server side, and places the more local-capable management jobs like replica created and queue maintained on the OSD side.(2) Proposed Agent technology in object storage system. Agent will create replica,maintain queue,modefy attribute,gather meterdata of object and so on. Agent can help to implement management of replica and advance efficiency of replica management.(3) Proposed a storage field based replica creating strategy, which avoids random replica placement within the system, but rather partitions storage nodes into different storage fields according to their bandwidth, available service and client situation. After evaluating the relativity within & between the fields, a replica can be placed at a proper node in a proper storage field.(4) Bring on the multi-dimensional replica management mechanism, which treats an object's key attributes or all ones as dimensions, to construct a multi-dimensional management view. In replica locating, we propose an expanded search tree, on which the optimized tree and one-path tree (an improved optimized tree) algorithm is realized. We proposed an optimized tree and the improved one searching algorithm. Current simulation manifests their satisfying performance.(5) Proposed a replica-based duplex load balancing strategy (DLBS) to better assess load and maintain balanced. In general, DLBS is a dynamic combination of an active strategy working on replica reproduction control, and another passive strategy that supervises hot-spot appearance, and provides transfer affairs if needed. They two together offer a model in handling load balance influenced by replicas. Simulation results and theoretical analysis demonstrate that DLBS can be utilized in the real OBS system to provide efficient load balance.(6) Useing the Petri Theory to characterize systemic load balancing. It is demonstrated that Petri Theory is both accurate and sufficient to describe control of the migration strategys. As a strong description tool, Petri is quite worthy as theoretical analysis in OBS.When we look back to previous research work and our current job, especially when we realize that memory capacity has no more been an obstacle for data storage, consuming a little capacity to introduce replica technology to Object-based Storage system is quite beneficial in that it will greatly enhance the overall performance.