RFLUSH: Rethink the Flush

Abstract

A FLUSH command has been used for decades to enforce persistence and ordering of updates in a storage device. The command forces all the data in the volatile buffer of the storage device to non-volatile media to achieve persistency. This lump-sum approach to flushing has two performance consequences. First, it slows down nonvolatile materialization of the writes that actually need to be made durable. Second, it deprives the writes that do not need to be made durable of an opportunity for absorbing future writes and coalescing. We attempt to characterize the problems of this semantic gap of flushing in storage devices and propose RFLUSH that allows a fine-grained control over nonvolatile materialization. The RFLUSH command delivers a range of logical block addresses (LBAs) that need to be flushed and thus enables the storage device to force only a subset of data in its buffer. We implemented this fine-grained flush command in a storage device using an open-source flash development platform and modified the F2FS file system to make use of the command in processing fsync requests as a case study. Performance evaluation using the prototype shows that the inclusion of RFLUSH improves the throughput by up to 6.5x; reduces the write traffic by up to 43%; and eliminates the long tail in the response time.