Understanding the performance of storage class memory file systems in the NUMA architecture

摘要

Recent developments in storage class memory (SCM) such as PCM, MRAM, resistive RAM (RRAM), and spin-transfer torque (STT)-RAM have strengthened their leadership as storage media for memory-based file systems. Traditional Linux memory-based file systems such as Ramfs and Tmpfs utilize the Linux page cache as a file system. These file systems have unnecessary overheads when adopted for SCM file system. Therefore, we propose a new memory-based file system using Memory Zone Partitioning called ZonFS, by extending the Linux Ramfs. In particular, we define a storage zone for SCM, modify the Ramfs to allocate a file system page from SCM. ZonFS avoids running Linux VM kernel codes such as (i) inserting pages allocated from SCM into the LRU list for VM page replacement and (ii) checking dirty pages for write-back to disk. Our extensive evaluations indicate that ZonFS has up to 9.1 and 14.1% higher I/O throughputs than native Ramfs and Tmpfs. Moreover, we also analyze performance behavior of ZonFS under the non-uniform memory access architecture of SCMs on a 40 manycore machine with various configurations such as file sharing level and file stripping level. Our evaluations show that memory controller contention and inter-node link congestion significantly affect the file system's performance and scalability.