Research On Cache Strategy For Distributed Object-Based File System

The Object-based Storage provides a cross-platform, high availability and highperformance storage architecture to meet the need of large-scale data storage. However,the object storage system will encounter I/O bottlenecks and the I/O performanceimprovement for storage device is limited. In this case, the cache technology developmentprovides effective solutions.According to the Cappella file system, designing a cache strategy on distributedobject file system, which implements a set of independent user mode caching scheme onOSD and achieves both optimization and improvement compared with the original one.The cache strategy improves the cache space allocation and recovery efficiency throughthe unified management and schedule of the cache space. It is also able to make anintelligent response to the cache resource usage and control of read and write strategies.When the cache resource usage reaches the upper limit, the cache strategy starts to persistwrite caches and eliminate read caches. On the contrary, when the cache resource usagereaches the lower limit, the cache strategy stops to eliminate read caches. The write cachespersistence operation will persist writing data to disk within the time window. The readcaches eliminate operation will begin to release cache resource with lower weightaccording to the weight of read cache.The write strategy achieves deferred write and aggregation of data, which effectivelyreduces the write latency and the number of disk writes. The read strategy optimizes dataread according to the characteristics of read request. It also achieves sequential read-aheadcache and read under partial cache hit optimization options. The cache strategy canachieve a better I/O performance by using direct I/O while occupying less memory usageand not relying on the Linux kernel page cache. Through optimizing relevant parametersspecific to real environment, the cache strategy can also be a good use in other distributedobject file systems.Test results show that the cache strategy achieves better optimization of the originalprogram, and has better performance than the Linux kernel page cache in the case ofsequential read and write for large files.