The Reliability Study Of Distributed Storage System

With the expanding of distributed storage system, the study of system reliability has drawn more and more people's attention. The increase in disk amount and the capacity of a single disk causes more and more errors in storage systems, anyone of which can cause great loss.In this dissertation, the system adopts erasure codes as its redundancy mechanism. Firstly, we calculate the probability of block errors using Poisson process and analyze the relationship between scrubbing interval and the probability of block errors; then we deduce the file availability model based on erasure codes and the probability of block errors, and we analyze the relationship between redundancy degree and file availability; finally, we establish the system reliability model based on file availability and scrubbing interval.In order to avoid accessing unrelated data blocks during read and write operations, we choose file as the redundancy group. Through analysis, we find that system reliability depends on the reliability of small files. So we propose the dynamic redundancy mechanism, which can choose the most appropriate redundancy degree for the files with different sizes. Dynamic redundancy mechanism not only reduce the space cost without affecting the system reliability, but also optimizes the performance of read and write operations to a certain extent.The common method to handle block errors is disk scrubbing, that is, the system scans all blocks in disks and repairs the corrupted blocks at intervals. So shortening the scrubbing interval can improve system reliability, however, over-frequent disk scrubbing will bring heavy burden on the system. At last, we will show how to choose an appropriate scrubbing interval based on the objective system reliability.