Modeling SMR Drive Performance

Shingled Magnetic Recording (SMR) provides higher track density by overlapping adjacent tracks, resulting in data loss on a “downstream” track if the upstream track is overwritten. To prevent such loss, disk surfaces are typically divided into bands separated by guard regions wider than the head width; these bands may then be sequentially written and re-written without cross-band interference. Drivemanaged SMR devices handle the resulting write constraints by incorporating an internal translation layer, much like SSDs, presenting a standard re-writeable block interface to the host. Generally-available SMR drives seen to date utilize a variant of the set-associative translation layer [2]: a static LBA-to-physical location mapping is used, writes are staged into a log or persistent cache, and are then de-staged to their “home” location via a band-at-a-time cleaning process. In this work, we present a model based on detailed evaluation of the Seagate ST5000AS0011, a 5 TB device-managed SMR drive, which predicts mean latency and throughput within less than 3% in most cases. It predicts tail latency, measured as RMS latency error, within 25%—only modestly higher than the 10% variation seen between nominally identical drives.