Study On Energy-Aware Prefetching Techniques For Disk Storage Systems

Since disk serves as the dominant storage device in data centers, conserving disk energy plays a significant role in data center operating fee reduction, and energy-saving and emission-reduction. However, implementing disk energy conserving technique is still facing many challenges currently. On the one hand, spinning up a disk consumes much power and time. On the other hand, disk has limited power cycles. Most previous work on energy-aware prefetching ignores the relationship between system performance and disk lifetime. Besides, the change of the state of a single disk may affect the performance of the whole storage system. Therefore, to address these critical issues, efforts must be paid to study disk energy conservation considering disk reliability and system performance, and the self-organizing behaviors in disk storage system.Power-aware Greedy Prefetching (PGP) increases disk idle intervals by greedily preloading much data to buffer. However, this might result in less energy conservation and more disk power cycles without exploiting the relationship between I/O access pattern and application pattern. Combining the tradeoff among disk power consumption, performance guarantee and disk reliability together, a power-aware prefetching framework for massive disk storage system, is proposed.To solve single disk and single stream case, a Disk characteristic based Power-Optimal Prefetching (DiscPOP) model, is formulated as an optimization problem. DiscPOP is proved to be solved via a 0-1 Integer Linear Programming (ILP) technique. For offline cases, a Greedy Partition algorithm (GP) is proposed to divide the problem into several small ones and solve them separately via the proposed ILP algorithm. For online cases, two heuristic algorithms are proposed based on Lazy Start Power-Optimal Prefetching (LSPOP) technique. Both of them use simple threshold controlled algorithms to select a prefetching start judiciously and cautiously. A Supple Chain Management based model is borrowed to solve online multi-stream power-aware prefetching. For multi-disk case, disks are divided into n groups to achieve n-competitive power-optimal prefetching. The results show both GP and online algorithms outperforms PGP and other traditional aggressive prefetching algorithms by more disk energy conservation and less power cycles.Conventional asynchronous prefetching incurs serious mixed random write and read load in SSD (Solid State Disk). To apply the mixed cache with SSD and DRAM (Dynamic Random Access Memory), three power-aware prefetching rules, are proposed. The prefetching degree and trigger distance of different streams are adjusted dynamically. SSD and DRAM serve different arrival ratio streams. When asynchronous prefetching is performed on SSD, some of the fetched data are placed in DRAM to eliminate mixed read and write load. Based on these three proposed rules, a Coordinated and Adaptive Prefetching (CAP) algorithm is proposed to improve multiple sequential prefetching in such hybrid caches. A page management mechanism is re-designed to reduce defragmentation in SSD. The results show CAP improve system throughput and reduce write allocations in SSD.An ideal power-aware data placement method for massive disk storage systems is proposed based on data access frequency. The replicas are grouped to reside in different disk groups to provide power-proportional storage service. Based on the ideal data placement, a 2-D cellular automata model, named MDSCA, is proposed to analyze and emulate the dynamical behavior rules in massive disk storage systems. The simulation results show that complex temporal and spatial phenomena evolve from the adjustment of local cells by energy-aware prefetchng and data migration. The total number of replicas increases when the load becomes heavier, and it tends to a stable sate eventually. Moreover, when the load is low, it is shown that there is an approximate power law distribution of the entropy of request queue length of each disk. To a certain extent, the whole system exhibits self-organization.