Minimizing Read Seeks for SMR Disk

Log-structured storage systems are file or block storage systems which write data in temporal order, rather using e.g. LBA or other spatial information to determine physical location. They are widely used as translation layers for flash and now Shingled Magnetic Recording (SMR) disk, performing the out-of-place write function needed for media where direct overwriting of data is not possible. Although certain aspects of log-structured storage system performance (i.e. write amplification) have been extensively studied, few or no efforts have examined the effect of seek overhead on workloads in this context. We examine seek overheads due to logstructured writing across a range of well-known and more recent block I/O traces, demonstrating that while some workloads are relatively unaffected, others suffer significantly. In addition we propose two novel mechanisms,opportunistic defragmentation and translation aware selective caching, for seek reduction in log-structured systems, as well as an application of existing disk read-ahead techniques (which we name translation aware look-ahead-behind prefetching), and demonstrate that these mechanisms greatly reduce or even eliminate increases in read seek cost in almost all workloads examined. We evaluate our proposed techniques using a metric we define, seek amplification factor. Results show up to 4x, 4x and 18x improvement of seek amplification factor among the studied workloads for opportunistic defragmentation, translation aware selective caching and translation aware selective caching respectively.