Trail: A Fast Synchronous Write Disk Subsystem Using Track-Based Logging

Synchronous disk write presents a serious performance problem for those applications that require strong data integrity guarantee. This paper describes the design and implementation of Trail, a disk subsystem that features a track-based logging technique to minimize the performance impact of synchronous disk write operations. We have demonstrated in this research that Trailcan reduce the latency of a synchronous disk write operation to close to data transfer delay only, the absolute theoretical minimum. Trailfirst buffers an incoming synchronous disk write request at a persistent logging disk, and asynchronously performs the write operation to a normal data disk. As a result, the performance of synchronous disk write is improved significantly without any compromise in data integrity. In the event of system failures, Trail’s data recovery mechanism guarantees the integrity of data blocks that are buffered at the log disk but not yet committed to the actual data disk at the time of failure. The current Trailimplementation is embedded within a SCSI device driver and is thus completely transparent to the rest of the kernel, including the file system. Measured performance results from the first operational Trailprototype under Linux show that the Trail-based file system is more than an order of magnitude faster in synchronous disk write latency than the standard Linux file system without Trail, while offering the same level of integrity guarantee. At the application level, the TPC-C performance of a Trail-based transaction processing system is on average 2.7 times better than one based on standard disk subsystem.