Exploiting The Parallelism Of Multiple Channels To Accelerate Persistent Memory File System Reads And Writes

Persistent memory file systems can realize efficient file read/write operations by leveraging Direct Access(DAX)that directly copies data between user buffer and persistent memory.However,analysis reveals that existing DAX-based file systems still use a singlethreaded and serial method in reading and writing each individual file.This fails to utilize the high bandwidth of persistent memory with multiple channels that can be accessed in parallel,thereby limiting the performance of file systems.To solve the problem,this paper proposes p Helper,a multi-channel parallelism-based approach to improve the file read/write performance of persistent memory file systems.Inside a file system,p Helper parallelizes the data block read/write process of individual file read/write operations by using extra concurrent threads(helper threads),so as to fully utilize the high bandwidth of multiple channels and hence improve read/write performance.p Helper mainly employs two strategies to improve file read/write performance from a local and global perspective,respectively.First,it uses a parallelism-aware request scheduling strategy to optimize the bandwidth utilization of each file read/write operation.It subdivides a file read/write operation into multiple concurrent persistent memory read/write requests adaptively to the number of available helper threads,and then schedules them among helper threads with load-balancing,thereby maximizing the bandwidth utilization of individual file read/write operations.Second,it employs a benefit-aware bandwidth allocation strategy to improve the total bandwidth utilization.It evaluates the benefit of using helper threads for current file read/write operation according to real-time bandwidth utilization,and the maximum bandwidth predicted by linear regression,so as to dynamically controls the number of helper threads,thereby achieving adequate and reasonable bandwidth utilization to persistent memory and improving the overall performance of the file system.p Helper is applied to two widely-used persistent memory file systems – PMFS and NOVA,and various workloads are used to evaluate their performance on Intel Optane persistent memory.The results show that p Helper can fully utilize system resources to increase the the performance of file read/write operations,and gain performance improvements by up to 8.9x and 2.5x for file read and write operations respectively.