Research On Shared Object-based Device Parallel File System

The development of cluster computing has made the problem of I/O bottleneck more and more difficult to deal with. To resolve the problem partially, researchers have begun to pay their attention to a new architecture for storage called as object storage architecture. In this thesis, a parallel file system is explored based on the architecture, which will contribute to providing the application of cluster computing with high performance, scalable file service and reliable data-storage. The details of results are as follows:1. A new method based on dynamical hashing partition is put forward to manage and organize the metadata. Under the condition of maintenance of uniform namespace, the method can achieve distributed management of metadata of file system by using a group of computers with shared storage, which will help to make the access to metadata more effective and more scalable and help to support fault- tolerance.2. An algorithm for data placement is proposed, which supports high availability. It, as a deterministic pseudo-random algorithm, expands the standard balls into bins model, and assigns data to different storage devices by partitioning of the unit range, which will support to assign the different redundant objects of one data object to different storage devices with the guarantee of the distribution of data with faithfulness and adaptation.3. A mechanism to fast recovery is provided. It can reduce the time spent to resume, when system fails, by execute the recovery of failed devices parallelly. And a model for distribution of mean time to data loss is built to estimate system reliability under different redundant mechanism.4. A protocol is put forward, which can support crash-recovery model, make seamless fault- tolerance, and achieve the Consistency of multiple replicas in asynchronous system. And a proof that the protocol maintains the linearizability is given.5. An applicable method to guarantee the atomicity of concurrent and overlapping I/O operation is provided, which is developed from custom file-locking and combined with the mechanism of token-passing.