Design And Implementation Of Workload-adaptive RAIS-5 System

With the rise of mobile Internet and cloud computing technology and the rapid development of mobile phones, sensors and the tremendous expansion of kinds of mobile terminals, boosting data exponential increase and leading us into the big data era naturally. Although opportunities brought with big data, challenges with respect to scalability, reliability, performance, of storage systems also appear. The widespread use of traditional RAID technology and flash-based solid-state disk(SSD) make the flash-based SSD array system a hot spot of current research, bring in hope to confront chal enges in big data area.Because of its simplicity and reliability as well as good parallelism, currently RAID-5 system has become the mainstream choice of diverse storage systems. However, the following problems exist in current RAID-5 systems:1. Only single type of redundancy policy is used;2. The existing redundancy policy is independent of system performance, making it unable to boost the system performance;3. The existing redundancy policy takes a unified approach towards all the workload data, which is lack of data-specificity;4. The existing redundancy policy place data through the round robin method, aiming to achieve load balance globally, but this method is not optimized for data access locality.On the other hand, extensive research and experiments indicate that most of the workloads take on an obvious feature of data access skew, meaning hot data, which accounts for a very small part, bear most access requests. In addition, through massive survey and analysis it is found that the storage of any single device in the array is usually not completely exhausted, always leaving some free space available.To this end, this paper propose a new design of a flash-based adaptive RAIS-5 system by making use of workload access skew and free space. In the ARAIS-5, hot data is effectively recognized and is then backed up by replication for once while cold data is striped as in common RAID-5 system. In this way, reliability of raw RAID-5 is preserved and by applying a local data placement for only hot data replica, data is evenly distributed across different SSDs. And with the help of newly created replica of hot data to alleviate most access pressure, better system load balance is achieved, thus enhancing whole system performance.In the implementation of ARAIS-5, we experiment with several real-world workload trace available in the Internet and results have proven the effectiveness in ARAIS-5.