基于预取的磁盘存储系统节能技术研究

发布时间：2018-04-02 10:13

本文选题：磁盘存储系统　切入点：预取　出处：《华中科技大学》2012年博士论文

【摘要】：磁盘占据数据中心数据存储的统治性地位,磁盘的节能控制对降低数据中心总运营成本和节能减排都具有重要意义。然而,磁盘的节能控制在实际系统的运用中却存在诸多困难和挑战。磁盘能耗状态的转换过程需要耗费较长的时间和较多的电能,容易造成系统读写服务的响应时间延迟,并且会影响磁盘的使用寿命,以往基于缓冲区预取的磁盘节能方法,大多未就系统性能和磁盘寿命进行综合考虑。另一方面,对单个磁盘的节能控制会影响数据中心存储系统的整体性能。研究保证磁盘可靠性和系统总体性能的能耗感知预取技术以及磁盘存储系统的自组织节能行为,是一项重要和紧迫的研究课题。围绕基于预取的磁盘存储系统节能技术,从理论分析、系统设计和实验测试得出了以下一些研究成果。现有的能耗感知贪婪式预取方法PGP (Power-aware Greedy Prefetching),通过将请求数据提前读入内存增大磁盘空闲时间间隔,是一种启发式的、具有一定实际效果的磁盘节能方法。在对PGP的预取机制分析中,发现PGP未对预取的启动时间和预取长度进行分析。进一步研究发现磁盘的空闲时间序列可能因预取启动时间和预取长度变化而变化,从而导致磁盘的总节电量减少和能耗状态转换次数增加。综合磁盘本身的属性、任务序列整体性能要求、磁盘的节能以及磁盘可靠性,建立了针对大规模数据中心磁盘存储系统的能耗感知预取优化框架。建立了针对单磁盘单数据流能耗感知预取优化模型DiscPOP,磁盘的节能目标函数和约束条件被证明为0-1整型线性规划问题。磁盘空闲时间序列过长,导致求解DiscPOP最优解的复杂度提高。贪婪分割算法是一种离线的、分而治之的策略,过滤掉磁盘空闲时间序列中的连续无效序列,将总的空闲时间序列分割成较短的子序列,通过线性规划解决器得到各个子序列的能耗感知预取最优解。提出了基于延缓开始机制的能耗感知预取在线算法,通过简单的控制条件,使得系统智能地选定一个启动点进行能耗感知预取,达到节能优化和磁盘能耗周期转换次数减少的目标。利用基于库存论的供应链管理模型,提出应用于多数据流的单磁盘能耗感知预取方法。通过对磁盘进行分组,利用单磁盘能耗感知预取最优解,提出一种2-竞争性的多磁盘能耗感知预取优化方法,并扩展至多组磁盘或者具有镜像磁盘的结构中。经实验验证,DiscPOP及其扩展方案降低了磁盘能耗并减少了磁盘能耗转换周期次数。研究固态盘和DRAM (Dynamic Random Access Memory.动态随机存储内存)组成混合缓存结构的能耗感知预取方法。通过对多顺序流的异步预取分析,发现混合缓存中的固态盘不仅会产生严重的写放大问题,还会产生严重的交织随机读写负载。提出了三个针对混合缓存的能耗感知预取优化规则,分别是通过对顺序流进行分类动态调整预取长度和触发距离、将不同到达速率的顺序流数据分别缓存在固态盘和DRAM上以及将固态盘上同一个顺序流的部分异步预取数据缓存于DRAM中消除交织读写情况。基于这三个预取规则提出了一种启发式的、面向混合缓存的协同式自适应能耗感知预取算法CAP,并重新设计缓存设备固态盘上的页面管理机制,降低固态盘作缓存时产生的碎片程度。经实验验证,CAP提高了系统的吞吐量,减少了固态盘写入速度,并优化了磁盘上的空闲空间序列,减少了磁盘的耗电量。提出了一种针对数据中心大规模磁盘存储系统的理想化能耗优化数据布局方法。依照数据访问的频度筛选出热点数据,并将其多个副本按照分组分别存储在各个磁盘组上,为大规模磁盘存储系统提供与能耗成比例的服务,需要打开的磁盘个数与需要提供的数据访问吞吐量成正比。利用动力学方法建立了一个针对大规模磁盘存储系统节能分析的二维元胞自动机模型。分析数据中心大规模磁盘存储系统的自组织性和自我调节能力,通过局部数据节点的能耗感知预取和数据迁移等行为,利用简单的状态转换规则,模拟和分析局部磁盘节能行为对系统整体性能和能耗的影响。实验结果表明,整个系统性能和节点状态随着局部磁盘的调控,元胞状态呈现出复杂的时空演化现象,副本个数随着负载的增加而增多并趋于稳定。在负载到达速度较低的情况下,各个磁盘的等待队列长度熵出现近似的幂律分布,整个系统的节能行为表现出一定的自组织特性。
[Abstract]:Disk data storage data center occupies the dominant position, energy saving control disk is of great significance to reduce data center total operating costs and energy saving and emission reduction. However, the energy saving control disk has many difficulties and challenges in the application in practical system. The conversion state of the disk energy consumption process takes longer and more power, easy to cause the system to read write service response time delay, and will affect the service life of the disk, the disk based energy saving method of buffer prefetching, mostly not on the system performance and the life of the disk into consideration. On the other hand, will affect the overall performance of the storage system of data center energy saving control of a single disk study on self organization behavior. To ensure such energy aware reliability and overall performance of pre fetching system disk and disk storage system, is an important and tight A compelled research topic.
The following research results are obtained from theoretical analysis, system design and experimental test based on the energy saving technology of disk storage system based on prefetching.
Energy aware greedy type pre existing method PGP (Power-aware Greedy Prefetching), the requested data is read into memory in advance increases disk idle time interval is a heuristic, disk energy saving method has a certain practical effect. The mechanism of PGP in the pre analysis, the PGP is not found on the pre fetching start time and prefetch length is analyzed. Further study found that free time series may be due to the disk prefetching starting time and prefetch length changes, resulting in the total electricity saving and reduce energy consumption of disk transition times increased. The attribute synthetic disk itself, the overall performance requirements of task sequence, disk and disk to establish such reliability. The energy aware pre for massive disk storage system data center optimization framework.
A single data stream for single disk energy aware prefetching optimization model DiscPOP, energy saving objective function and constraint conditions of disk prove to be 0-1 integer linear programming problem. Disk idle time series is too long, lead to increased complexity of solving the DiscPOP optimal solution. The greedy segmentation algorithm is an off-line, divide and rule strategy, filtering continuous off invalid sequence disk idle time sequence, the total idle time sequence is divided into short sub sequences by linear programming to solve energy aware for each sub sequence of the pre take optimal solution. The proposed energy aware mechanism for online pre start delay algorithm based on the control condition is simple, which makes the system intelligent selection of a starting point for energy aware prefetching, achieve energy optimization and energy conversion to reduce the number of disk cycle target. Using the theory of stock based on Supply Chain The management model, proposed a single disk energy aware pre applied to multiple data streams is proposed. By grouping the disk, get the optimal solution by a single disk energy aware pre, proposed a multi disk energy aware pre 2- competitive optimization method, and set up the extended structure disk or disk image. After experimental verification, DiscPOP and its extension scheme reduces energy consumption and reduce the energy consumption of the disk disk conversion cycles.
Study on the solid state disk and DRAM (Dynamic Random Access Memory. dynamic random access memory) composed of energy aware pre mixed buffer structure is proposed. By the pre asynchronous order flow analysis, found in the mixed solid state disk cache will not only write amplification problem is serious, but also have a serious interleaving random read and write load. This three for hybrid cache energy aware prefetching rules, are classified to dynamically adjust the prefetch length and distance based on the trigger sequence flow, different arrival rate of order flow data were cached in solid state disk and DRAM and asynchronous pre solid state disk with a sequence of data flow the cache in DRAM to eliminate interleaving read and write cases. The three prefetching rules proposed a heuristic algorithm based on CAP collaborative adaptive energy aware prefetching for hybrid cache, and The design of solid state disk cache device page management mechanism, the degree of fragmentation is reduced as generated by the solid state disk cache. Experiments show that CAP improves the throughput of the system, reduce the solid state disk write speed, and the optimal sequence of free space on the disk, reduces the power consumption of the disk.
The optimization of data layout method is ideal for large-scale data center energy consumption of a disk storage system. In accordance with the frequency of data access the selected hot data, and the multiple copies are stored in each packet according to a disk group, providing proportional service and energy consumption for large scale storage system, need to open the number of disks with the need to provide data access throughput is proportional. Using the dynamic method to establish a two-dimensional cellular automaton model for large scale energy saving disk storage system analysis. Analysis of self-organization and self adjustment ability of data center scale disk storage system, the energy aware node pre local data and data migration. Using simple state transition rules, impact simulation and analysis of local disk energy saving behavior of the overall system performance and energy consumption. The experimental results Show that the performance of the whole system and node state with the regulation of local disk, the cellular state showing the evolution of complex spatial phenomena, a number of copies as the load increases and tends to be stable. When the load reaches a lower speed, each disk queue length entropy shows power-law distribution approximation, energy saving the behavior of the system exhibits self-organizing characteristic.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：TP333.35

【参考文献】