Caveat-Scriptor: Write Anywhere Shingled DisksCarnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-15-101. April 2015.
Saurabh Kadekodi, Swapnil Pimpale, Garth Gibson
Carnegie Mellon University
Magnetic disks, under pressure from solid state flash storage, are seeking to accelerate the rate that they lower price per stored bit. Magnetic recording technologists have begun to pack tracks so closely that writing one track cannot avoid disturbing the information stored in adjacent tracks. Specifically, the downstream track will be at least partially overwritten, or shingled by each write, and the upstream track will tolerate only a limited number of adjacent writes. Some data that was stored in the downstream track will be lost, forcing firmware or software to ensure that there was no necessary data in parts of that track.
In order to avoid deployment obstacles inherent in asking host software to change before shingled disks are sold, the current generation of shingled disks follow the model established by flash storage: a shingled translation layer of firmware in the disk remaps data writes to empty tracks and cleans (read, move, write) fragmented regions to create empty tracks. Known as Drive-Managed Shingled Disks, host software does not need to change because the disk will do extra work to cope with any write pattern that could lose data. To reduce or eliminate this extra work, changes in the hard disk API have been proposed to enable Host-Managed management of shingled disks.
This paper explores two models for Host-Managed Shingled Disk operation. The first, Strict-Append, breaks the disk into fixed sized bands and compels disk writes to occur strictly sequentially in each band, allowing only per-band-truncate-to-empty commands to recover space. This is approximately a physical realization of the classic Log-Structured File System (LFS), and shares the need for the file system to schedule and execute cleaning of bands. The second model, Caveat-Scriptor, exposes a traditional disk address space and a few shingled disks parameters: a distance in the downstream block address space that is guaranteed to never experience shingled overwrite data loss and a distance in the upstream block address space that cannot tolerate multiple adjacent writes. Host-Managed software for Caveat-Scriptor shingled disks is allowed to write anywhere, but if it fails to respect these distance parameters, it may lose data. We show in this paper that Caveat-Scriptor enables reuse of previously written and deleted data with far less cleaning than Strict-Append, enabling the potential for high-density Shingled Disks to perform almost as well as lower-density non-Shingled Disks.
KEYWORDS: Shingled Disks, Filesystem, Log-Structured, SMR, Segment Cleaning, Aging
FULL TR: pdf