Concurrency-oriented Space Management Scheme For In-memory File Systems

In recent years,with the rapid development of cloud computing,artificial intelligence and 5G technology,the amount of data generated is increasing dramatically all the time,which puts great pressure on storage and calculation.The emerging non-volatile memory NVM is one of the mainstream solutions for the next generation of high-speed persistent storage.It has good characteristics such as access speed close to memory,no loss of data when powered down,byte addressing and low power consumption.Therefore,there are a lot of researches that use non-volatile memory as the bottom storage of file systems,and then many file systems designed for non-volatile memory are born.These file systems make full use of the characteristics of non-volatile memory and can achieve good performance.However,the existing non-volatile memory file system does not optimize the high concurrency environment and has poor performance scalability in the high concurrency environment.In the aspect of wear equalization,the existing NVM file system does not fully consider the problem of wear equalization among NVMs in systems with multiple NVMs,resulting in some NVMs in the system being more easily worn out,thus affecting the reliability and life of the system.In view of the above opportunities and challenges,this thesis proposes a space management strategy for wear awareness of non-volatile memory file systems in high concurrency environments.It improves performance in high concurrency environments by allocating private allocation areas to threads to avoid competition for file system page allocators.At the same time,it distributes read-write pressure to multiple NVM through scheduling strategy to achieve a combination of load balancing and wear balancing.In order to improve the performance in high concurrency environment,the strategy divides NVM space into thread private NVM space and file system public NVM space.Each thread has a private allocation area to avoid competition for file system page allocation area.This thesis proposes three different page recycling strategies to adapt to different system environments to achieve the highest efficiency.The page recycling strategy based on free space,the least use recycling strategy based on timestamp,and the least use recycling strategy based on direct mapping and the doubly linked list are proposed.This thesis implements the space management strategy on the existing non-volatile memory file system and makes experimental comparison.Experimental results show that,in terms of performance,the space management strategy can increase IOPS by up to 15% and write throughput by up to 7.6%.In terms of wear balance,the space management strategy can obviously improve the service life of NVM in the system.Through theoretical analysis,we get the decision table according to the size of NVM space for several page recycling strategies to adapt to various system environments and achieve the highest efficiency.