Research On Performace Optimization Of NVM-based Storage Systems

NVM is a next-generation storage material whose read and write speed is close to traditional DRAM.It can durably keep data without any extra energy supply and has high storage density,just like classical HDD and SSD.Due to these advantages,NVM has received a lot of favors from various people and organizations.However,as an emerging material,the related theories of NVM have not reached an acceptable level.The industrial skills cannot support a mass production on NVM.Moreover,the deeper the research and development on NVM,the more shortages come to our sight.If we directly deploy NVM into current storage system,not only we may not active the performance improvement,but also do negative influence on system,because of the improper usage of NVM.In order to solve these problems,this thesis working on optimization at NVM based computing and storage environment.By adapting NVM on serval popular applications,this thesis explores the proper usage of NVM at current storage architecture.First,we use NVM to optimize the popular hybrid storage architecture.By using the excellent features of NVM,we improve the performance of hybrid storage system.In order to fully utilize every NVM storage units,we modify traditional storage management mechanism and design an extra unit called “Data Bucket”.Based on data read and write characteristics,historical access information,etc.this thesis keep data on NVM according to NVM storage unit's remaining lifetime.Second,this thesis discusses several key factors affecting the persistence performance of distributed caching system for the persistence of distributed caching system.Based on the analysis of the key factors affecting the persistence performance,some pertinent changes have been made to the current persistence mechanism so that the NVM-based distributed caching system can exert its advantages better.Finally,this thesis uses the popular Spark platform as an example to introduce a shared cache structure based on NVM design into the hardware system of the Spark platform.By modifying Spark's persistence mechanism and RDD partition management algorithm,Spark can correctly identify and use NVM.At the same time,the NVM shared cache structure improves the Spark task processing capacity and the utilization of storage space.This thesis explores the reasonable application of NVM in the current storage-computing environment,through the combination of several common application scenarios and NVM.By proper modification of the current mechanisms,the performance and advantages of NVM are enhanced under current application environment,and some disadvantages of NVM are avoided.This thesis provide some meaningful reference for further application of NVM.