Research On Some Energy-Aware Techniques Of Flash Memory Storage System For Embedded Systems

Flash memory has been widely used in embedded systems as secondary storage media, because of its non-volatile, shock-resistant, and power-economic advantages. But the traditional designs of flash memory storage systems cannot satisfy large scale, high-performance, low power and reconfigurable requirements of modern systems. Our work focuses on the key techniques of energy-aware design and implementation of flash memory storage systems.First, the architecture of traditional flash memory storage system is thoroughly investigated and analyzed. The adaptive, reconfigurable and energy-oriented framework is proposed based on the separation management of space and energy.Second, two basic energy-aware strategies are applied to the flash memory storage system. On the one hand, based on the formalized concepts of exact data affinity as well as fuzzy data affinity, four key techniques are proposed to resolve the problem of applying dynamic power management under multi-bank architecture. The detection mechanism of fuzzy data affinity under multi-hash function framework combined with counter list is proposed. Recovery time is used to quantify the relation between energy state and bank allocation strategy. The product of recovery time and erase count is working as the allocation criteria. Transition of energy state is implemented with predictive single parameter feedback method. And the delayed write technique is proposed. On the other hand, the real-time dynamic-voltage-scaling algorithm reducing energy consumption is proposed.Third, extend the function of space management under the traditional architecture to energy-oriented optimization in two levels. The first level focuses on the energy-aware design of space management mechanism. Energy-oriented address mapping mechanism, called the framework of adaptive multi-log block hybrid address mapping is proposed. The concept of "property of data" is defined as the center of key techniques. Then the high-performance segmented LRU algorithm under multi-hash function framework of hot data identification mechanism is studies and implemented. The performance of hot data update is improved through dynamic allocation of multiple log blocks.The second level focuses on the framework of multi-strategy garbage collection. Energy power function quantifying the relation between energy states and recycling cost is combined with migration power function to be the selection criteria. EID recycling selective algorithm is proposed to provide the comparable wear-levellingdegree with cost-benefit strategy and ensure the energy optimization under multi-log block structure.Based on the above studies in this thesis, we implement three kinds of simulators: multi-bank architecture simulator, multi-log adaptive hybrid mapping simulator and dynamic voltage scaling of multi-task set simulator. The results of experiments show the effectiveness on energy reduction of the energy-oriented system framework based on the separation management of space and energy, the multi-bank dynamic power management based on detection of fuzzy data affinity, real time dynamic voltage scaling algorithm and garbage collection strategies.