Research On Energy-saving Data Layout And Performance Optimization For Disk Arrays

With the rapid development of IT technology and mobile Internet, the amount of data storage is growing rapidly, and with the increase of storage devices, energy consumption of storage devices has reached 26%~40% of the total energy consumption in data center. It is important to study the energy-saving of storage devices.Sequential data storage applications are widely deployed in recent years, such as video surveillance, continuous data protection, and virtual tape library, backup and archiving, Sequential data storage has inherent storage characteristics and access pattern, mainly performs sequential data access and requires lower random performance; has no rigorous requirement on bandwidth, but requires higher security and huge capacity; mainly implements writes to storage. These factors should be considered during the research of energy-saving sequential data storage system. We need to have suitability research on organizational structure of RAID, energy-saving mechanism and scheduling algorithm to exert the potential of storage devices and reduce energy consumption of the storage devices.S-RAID is an energy-saving disk array which suits for sequential data storage. S-RAID saves energy by dividing disks in the array into groups, and only a few of them run in parallel. The S-RAID balances the performance and energy consumption of the storage system, and reduces energy consumption of the storage devices under the conditions of performance requirements.In this thesis, the energy-saving RAID for sequential data storage system based on S-RAID is reviewed, the main contributions include:(1) The partial-parallelism data layout in S-RAID is static, which is suitable for smooth workloads, and difficult to adapt to fluctuate or burst workloads. A dynamic energy-efficient data layout(DEEDL) is proposed for above applications. DEEDL inherits the partial-parallelism energy-saving stratagem, and further adopts dynamic address mapping mechanism to allocate storage space with appropriate parallelism according to the performance requirements. DEEDL has prominent ability to adapt to fluctuate or burst workloads for continuous data storage. The results of simulation experiment shows that DEEDL is more efficient on energy-saving than the mode of S-RAID、PARAID and eRAID 5. Its’ 24-hour power consumption is only 83%, 29%, 31% of those of S-RAID, PARAID, and eRAID 5.(2) In order to make more disks into standby to saving energy, S-RAID will executive the “small write” operation called “read-modify-write”. Because “small write” lead to extra read operations during the writing, the write performance of each dish is low. Aiming at this defect of S-RAID, propose the method of performance optimization base on pre-reading and I/O polymerization to S-RAID. To improve the disk’s usage ratio in the way of reduce the number of I/O operations and seek data, increase the I/O operations size. Specific measures include: stable recognize write request stream from upper application; trigger the large size asynchronous pre-reading, old data and parity data required by pre-reading the small write operations by write request stream; write polymerization and merge some write requests into one or a few large-sized write request; build the write assembly line based on pre-reading, write cache and write back. These methods take advantage of the storage characteristics of sequential data storage applications and performance advantages of modern disks, improved the write performance of S-RAID significantly. With experiment of 98% the proportion of sequential write, it can improve the performance at least 47%, while 56% performance improvement in alternating write sequential flow.(3) In sequential data storage applications, it mainly to sequential access and also a few random accesses, but the random access may lead to low disk performance, on the other hand, the serious problem caused by “small write” may lead to low write performance. A high-performance and energy-efficient RAID called Ripple-RAID is proposed for these applications based on S-RAID. A new partial-parallel data layout is presented, and by comprehensively employing the strategies, such as address mapping based on SSD, out-place update, generating parity data progressively based on pipeline, cache optimize etc. The write performance and energy consumption of Ripple-RAID are better than S-RAID 5 while providing single disk fault tolerance. When write workload is 80% sequential and transfer request size is 512 KB, Ripple-RAID is 3.9 times the write performance of S-RAID 5, 1.9 times that of Hibernator and MAID, 0.49 times that of PARAID and eRAID 5, while conserving 20% energy than S-RAID 5, 33% energy than Hibernator and MAID, 70% energy than eRAID 5, and 72% energy than PARAID.