Research On Encoding For Optimizing Lifetime And Energy Of Non-Volatile Main Memory System

With the rapid development of big data technology,the demand for a large-size,low power as well as fast main memory system is more important than ever.Emerging Non-Volatile Memories(NVMs)are promising to replace the DRAM technology due to the advantages of high density,low static energy consumption and fast read speed.NVMs,represented by Phase Change Memory(PCM),can only endure a limit number of bit flips,and bit flips consume high energy.Therefore,NVMs face the disadvantages of limited lifetime and high write energy.Reducing the number of bit flips,improving the lifetime and reducing the write energy consumption are the focuses of the research about NVM-based main memory.This thesis reduces the bit flips of NVM by efficient data encoding techniques,thereby improving the lifetime and decreasing the write energy.According to the storage size of the encoded data,encoding methods can be divided into two categories,i.e.,encoding with tag bits and compression-based encoding.This thesis optimize these two methods to solve the problems of low utilization ratio and high storage overhead of the tag bits in encoding with tag bits,and similarity destruction in compression-based encoding techniques.Aiming at the problem that the existing encoding with tag bits has low utilization ratio of the tag bits due to ignoring the data patterns of the workloads,adaptive granularity encoding is proposed to further reduce the writes to NVM,improve the lifetime and decrease the write energy.In the methods of encoding with tag bits,every N-bit data is assigned a 1-bit tag.N is the encoding granularity.The efficiency of encoding will increase when N decreases.We observe that there are a large number of redundant words.This makes the utilization ratio of the tag bits of data encoding methods become very low.To fully exploit the tag bits to reduce the bit flips of NVMs,adaptive granularity encoding compares the old and new data to identify the redundant words,dynamically assigns the tag bits to the modified words and adaptively sets the encoding granularity according to the number of the modified words,thereby reducing the encoding granularity and the bit flips.We further observe that there are sequential flips in the workloads,and smaller encoding granularity leads to more bit flips of the tag bits in this case.To reduce the bit flips of the tag bits,based on “sequential flips”,adaptive granularity encoding dynamically selects the encoding granularity which causes the minimum bit flips instead of using the minimum encoding granularity.Experimental results show that adaptive granularity encoding can decrease the bit flips by 7.2%,reduce the energy consumption by 17.0%,and improve the lifetime by 3.7% compared with the state-of-the-art CAFO.The capacity overhead of the proposed scheme is 8.2%.Leveraging the advantage of size reduction of compression,SELective Encoding and Compression(SELEC)is proposed to reduce the bit flips,improve the lifetime and decrease the write energy with low capacity overhead.To significantly reduce the bit flips,the capac-ity overhead of the tag bits becomes unacceptable.Compression can reduce the size of the data to store and save the space.SELEC exploits the space saved by compression to store the tag bits of encoding methods,and therefore it can reduce the bit flips with negligible capacity overhead.The access granularity of main memory is a cache line.Furthermore,the saved space size of each compressed cache line varies,and different encoding methods have different tradeoffs between capacity overhead and effect.To fully exploit the space saved by compression for improving lifetime,SELEC selects the proper encoding method which can fully use the saved space on the condition that no extra space is used.To provide more space for efficient encoding,SELEC selects an efficient compression scheme from Fre-quent Pattern Compression(FPC)and Base-Delta-Immediate(BDI)algorithms.Still,some data patterns cannot be compressed by any compression techniques.We use the encoding method with low capacity overhead to encode the uncompressible cache lines.Experimen-tal results show that the proposed scheme can reduce the bit flips by 13.3%,decrease the energy consumption by 10.1% and improve the lifetime by 29.8% with 3.5% capacity over-head,compared with the Flip-N-Write with 12.5% capacity overhead.Encrypted NVM suffers from heavy writes due to “avalanche effect”.Existing encod-ing with tag bits can already be used to reduce the bit flips of encrypted NVMs.However,compression cannot be applied to reduce the bit flips of encrypted NVMs.Based on the observation that employing compression directly in encrypted NVM destroys the “data sim-ilarity” and leads to more bit flips,Similarity-Aware Compression Scheme(SACS)is pro-posed.The unmodified data between the old and new cache line cause the “data similarity”.Compression changes the value and location of the data to be written,and therefore the “data similarity” is destroyed.SACS uses compression algorithm to decide the size and location of the data to write and writes the compressed data to their locations before compression.Leveraging the size reduction of compression,SACS enables fine-grained dirtiness track-ing,finds more unmodified data and reduces the data amount to be re-encrypted and written by adapting the tag bits of recording dirtiness to the compressed data.When writing the compressed data to NVM chips,the write latency is restricted by the slowing chip due to the non-uniform byte-level writes.SACS distributes the bytes of compressed data evenly to the NVM chips and reduces the write latency through reorganizing the data layout of mapping cache lines to NVM chips.Experimental results show that SACS improves the lifetime by27%,reduces the write energy by 24%,decreases the write/read latency by 50%/70%,and improves the IPC performance of encrypted NVM by 11% compared with the state-of-the-art DEUCE.The capacity overhead of SACS is 8.8%.