Research On Energy Consumption Management Strategy For Distributed Storage System

Distributed storage technology with high storage bandwidth, good scalability and manyother features is the mainstream direction of technical of mass storage system for meeting therequirement of the explosive growth of mass data storage. However, the electrical energyconsumption of distributed storage system has become the largest expenses in the course of itsoperation. In addition, the consumption of a large amount of electrical energy indirectlyincreases CO2emissions, which will exacerbate the global greenhouse effect. As aconsequence, the issue of saving the energy consumed in the distributed storage system hasbecome a hot research problem in computer storage field. Meanwhile, it is a very significantresearch topic both in theory and practicality, and has important significance on the followingthree aspects:(1) reducing the energy consumption in distributed storage system;(2) savingstorage cost; and (3) being in line with national energy conservation policy.Energy consumption management strategy is an effective energy saving technology ofreducing the consumed energy in the distributed storage system, but these problems, such asthe traditional stochastic energy management strategy is not suitable for the dynamicapplication environments, the memory energy-consumption management strategy has greatinfluence on the performance of the storage system I/O, and the cooling subsystem of thestorage system has low energy-consumption management efficiency, still are the deficienciesin it. Therefore, the conducted research in this thesis mainly focuses on these three major issues,the dynamic load adaptive energy-consumption management strategy, the memoryenergy-consumption management strategy, and the energy-consumption optimizationmanagement strategy of the cooling subsystem. The innovative works and contributions ofthesis are stated as follows.1) A dynamic adaptive energy management strategy base on iterative parametersestimation (IPE-DAEM) is proposed, which aims at the poor adaptability of the traditionalstochastic energy management strategy on dynamic load adaptability, firstly, the randomiterative parameter estimation method is used to estimate the load parameters, and then aglobal static Markov control model is established by combining the component status.Secondly, the constraint equations meeting the performance and energy-consumption arededuced. Finally, the optimal energy-consumption management strategy can be obtained bysolving the linear programming. The iterative parameter estimation method can quickly andaccurately get the status parameters of the dynamic load with little computational cost, which sets the stage for the application of static Markov model to dynamic environment. Then wecan effectively apply the calculation accuracy of the static Markov model into the dynamicload. For dynamic load problem, IPE-DAEM shows a better adaptability compared to thetraditional stochastic energy management strategy based sliding window parametersestimation..2) A memory energy management strategy base on replica mechanism (RM-MEM) isproposed, which aims at the performance delay caused by the energy-consumptionmanagement strategy, and the inefficiency on I/O performance in memory energymanagement, the replication space is established in the memory by using the programmableNIC to improve on the existing memory energy management. In RM-MEM, thehigh-frequency read-only data can be stored in the NIC replication space, and a replicationmetadata list built in the memory guides the host computer to use the replicates. In addition, areplacement algorithm based ratio-temporal is designed to ensure the replication validity, andthe commands keeping the replication metadata in active state should be added into thehigh-performance energy-consumption management algorithm without changing the othercontents. Therefore, high-frequency read-only data are no longer repeatedly transmittedthrough the Bus in the I/O process, which reduces the Bus traffic and further enhances thesystem throughput. The theoretical analysis and experimental results indicate that the NICreplication of the proposed new strategy filters the high-frequency read-only data in the I/Oprocess, increases the memory free time, effectively improves the utilization of the Busbandwidth, reduces the I/O time, and consequently overcomes the traditional strategies’negative effect on the I/O performance with obviously reduced energy consumption.3) A cooling subsystem energy consumption management strategy base on temperatureequilibrium (TE-CSEM) is proposed. For the existing cooling strategies used in currentdistributed storage cooling sub-system, the passive adaptation to the temperature field has anunsatisfied performance, the feedback operation performed in the active temperature balancemechanism may leads the kernel cooling redundancy, and the prediction based hot spotselimination approach may affect the performance. After deeply analyzing the thermaldissipation mechanism, we presented a multi-level-task-scheduling cooling optimizationstrategy in which the thermal load balancing among the CPUs is achieved by reactivelyallocating the corresponding load to the CPUs. If there was a overheating kernel, the hot spotswould be eliminated by CPU thread scheduling. Meanwhile, the smooth regression predictionmodel is used to accurately predict the temperature in inner CPU, and then, a kernel threadscheduling algorithm based on a predetermined temperature is designed to further eliminate the hotspot. The theoretical analysis and experiments indicate that the proposed methodovercomes the disadvantage of unsatisfied cooling efficiency of traditional coolingenergy-consumption management without influencing the system performance. The reasonsare as follows: firstly, the balanced temperature makes the cooling fan keeping in minimumaverage rotation speed and consequently minimizes the energy consumption; secondly, thetime caused by the thread-level scheduling can be ignored compared to that caused by theOS-level tasks scheduling.4) A middleware based energy-consumption management application framework isproposed, which can integrate all the energy-consumption management strategies and locatesbetween the operating system and user application layer The proposed framework comprisestwo parts, i.e., the component-level energy-consumption optimization part and theequipment-level energy-consumption optimization part. For different storage tasks withdifferent performance requirements, the framework restrictively provides the minimumserving nodes and shut down the inactivity nodes to save the energy. For the componentsbelong to these active nodes, the component-level modules will assign suitableenergy-consumption management strategy for active node’s components to further reduce theenergy consumption. The proposed energy-consumption management application frameworkcan meet the service requirements put forwarded by different users and effectively reducesystem energy consumption by using the energy consumption management strategy.