Research And Design Of Writing Strategy Optimization Based On NVM

In order to alleviate the speed difference between CPU to the memory, to the external memory, the existing storage infrastructure has continued the initial layered architecture:cache, memory and external memory, from high layer to low layer and the speed from fast to slow, however, with the application of massive data storage and large scale parallel computing, more and more extensive applications, memory performance needs to be improved. Whether external memory or memory from the storage structure to the storage medium are faced with the challenge of emerging technologies. To the external memory aspect, SSD(Solid State Disk) used NandFlash as a storage medium have made a good record and the research on relative technology is still going on and the performance of SSD is becoming better; to memory, constant research in the field of production technology to enhance the DRAM storage density, in recent years, the effect is poor, the academic field will turn to look for new storage media to replace the traditional DRAM. NVM(Non-Volatile Memory, like PCM, STT-RAM, FeRAM) has many attractive properties, such as power loss, high density, high impact resistance, low energy consumption and excellent scalability. In addition, NVM has a very low free energy consumption compared to DRAM and higher storage density. Therefore, NVM is becoming more and more popular, in numerous systems have been adopted. It can not only as a secondary storage, can also be as the main memory of the embedded system, and there are a large number of techniques have been proposed using NVM replace traditional dram as the computer’s main memoryBut NVM itself also has a two obvious flaws:1) NVM write/erase cycles limited number (for example, phase change memory PCM cycle number only 109); 2) NVM need longer write operation time and higher write operation energy consumption (for example, PCM read operation time is 50ns, and write operation time of 1s). If we can find a solution to these problems, the promotion of NVM will play a crucial impact.In this paper, we propose the MinFS, a NVM write scheme with bit flipping and cyclic shifting, to minimize update bits. Similar to Flip-N-Write and Min-Shift, every PCM write unit has an extra flip bit that marks whether the associated data have been flipped or not and several shift bits by that the data have been cyclically shifted right. When a write request arrives, the MinFS first computes all possible encoded new data and reads out old data. Then, the MinFS computes all hamming distances between old data and each possible encoded data, in which all bits including data bits, flit bit and shift bits are used. The encoded data with the minimal distance is chosen as the optimal result and written into NVM cells. This scheme can find the minimal hamming distance between the data to be written and the original data and reduce the number of bit to write to the maximum extent. The MinFS can find the minimal encoded data at the cost of read, computation and comparison.We developed a trace-driven tool and evaluated our proposed scheme using random data and benchmarks. Experimental results showed that the MinFS scheme reduces average update bits by 8.2% and 5.7% over Flip-N-Write and Min-Shift respectively.This paper proposed a strategy from the perspective of coding method optimization to extend the life of NVM which has important practical significance for the promotion of NVM and have a reference significance for future theoretical research in related fields.