Optimization Of Wear-leveling-aware Multi-grained Allocation Mechanism For Persistent Memory File Systems

The demanded storage scale of upper-level applications is getting larger and larger driven by big data.Performance and power bottlenecks of traditional storage systems are more serious,becoming an important factor hindering the rapid development of upper-layer applications.In recent years,many new types of Persistent Memory(PM)with the advantages of byte addressing,non-volatility,low latency,low power consumption,high storage density,easy expansion,and vibration resistance have emerged.PM provides novel resolutions for the above bottlenecks and promotes the development of persistent memory file systems.However,these persistent memories commonly suffer disadvantages of low write tolerance,which seriously threatens the data reliability of the file system and brings new challenges to researchers.Existing wear-aware allocators based on persistent memory file systems only support single-granularity allocation,resulting in serious performance overhead,and will greatly reduce the performance of persistent memory file systems.To solve the above problems,this thesis research the optimization of the space allocation mechanism of the persistent memory file system considering the write wear,and proposed a Wear-leveling-aware Multi-grained Allocator(WMAlloc).WMAlloc achieves the wear-leveling of PM while improving the performance for Persistent Memory File Systems.The main contributions of this work are as follows.First,the Multi-Grained Allocating Heaps strategy(MGAH).MGAH uses multiple min-heap trees to divide the persistent memory storage space into blocks of multiple granularities.A min-heap tree manages blocks of the same size according to the degree of wear from low to high.Moreover,MGAH allocates the needed blocks from the appropriate min-heap root node based on the requested size.MGAH can achieve space multi-grained allocation and wear leveling of persistent memory.Second,the Wear-Aware Recycle Forest mechanism(WARF).To improve the recycling efficiency of persistent memory storage space,this work uses multiple red-black trees to manage the recycled PM space.Among them,different red-black trees manage recycling blocks with different wear ranges and sizes.In addition,WARF takes free blocks from the black tree with the smallest wear range and inserts them into the corresponding minimum heap each time.Third,the Wear-Balancing Node Migration algorithm(WBNM).To address the problem of extremely unbalanced min-heap node wear between the min-heaps caused by the imbalance of the application allocate block granularity,WBNM dynamically migrates the extremely worn nodes in the min-heaps.Finally,the typical persistent memory file system NOVA was modified in the Linux kernel to realize the proposed WMAlloc.This thesis compares WMAlloc with the existing state-of-the-art and wear-leveling-aware space management technique.The experimental results show that the lifetime of PM and performance in WMAlloc are1.48 times and 1.44 times that of DWARM,respectively.