The NVM technique advanced rapidly in recently years and has impacted our daily life in very corner. The smartphone, which uses NVM to storage data, has changed the way of how we use computers, while the AFF array storage system, which uses NVM to largely reduce the latency, have enabled the internet service providers to use machine learning method in a big data environment to improve their quality and accuracy of their service. Although each of them works in dierent scenarios, has dierent users, and has dierent goals, they still share lots of common insights and design experience when they use the non volatile memory (NVM) as the primary storage media. Fully understanding the infrastructures, key algorithms, and key limitations, and I/O stack of them can largely inspires vision and help discover superior insight and achieve unparalleled results in future research. In addition, comprehensively comparing the dierence between them also helps us to enhance our knowledge and helps us understand the real world problem in depth. In this research, we conduct an project on I/O characteristic analysis on smartphone based on real trace collected with a Nexus 5. Several insights are provided and two case studies are conducted based on simulation to verify the eectiveness of our insights. Then, a new le system call MSUFS is designed and implemented as the rst step toward the goal of operating system redesign for pure persisting memory based systems. Results show that we can achieve even better performance while provide enhanced data protection mechanism. Third, an I/O request scheduling optimization project has been done to improve the latency of WAFL based data storage systems. Results show that NVM based tuning for scheduling can obviously improve overall latency of a system. Followed this a computational study has been conducted. In this study, we propose a pieceswise function model to simulate the latency-throughput curve, which is use to estimate the overall performance of a WAFL based system. results show that the variation is minor and the method is able to reduce the necessary data points ,which further reduces the cost and time consuming of enterprise level system performance evaluation.