Fine-Grained Data Committing for Persistent Memory

摘要

Non-Volatile Memory (NVM) is better than traditional DRAM with respect to energy efficiency and larger capacity, so NVM has begun to be used as main memory. NVM provides data persistence that data written into NVM will not be lost during unexpected system failure occurs. Data persistence is mandatory for programs such as file system and database. However, traditional memory protocol cannot provide an mechanism for programs to guarantee data persistence because the write instructions do not ensure that data would be eventually written into the memory media. Furthermore, extra global operations such as PCOMMIT for data committing could incur significant performance loss, especially for multi-task programs. To address this issue, we propose a hardware-software coordinated mechanism to achieve low-overhead data committing. Write queues in memory controller are divided into multiple sub-queues for monitoring write commands for different address ranges. Programs can query write queues to check the execution status of previous written commands through a series of OS-managed library APIs. Fine-grained data committing can reduce the interferences among processes effectively. Extensive evaluations show that per-task data committing brings an average 1.78x performance improvement than original global committing mechanism and accelerates the data committing by 2.07 times.