The layout and replication techniques required to enable graceful degradation introduce a host of design issues. We highlight the major challenges that arise.
Semantically-related blocks: With fault-isolated data placement, D-GRAID places a logical unit of file system data (e.g., a file) within a fault-isolated container (e.g., a disk). Which blocks D-GRAID considers ``related'' thus determines which data remains available under failure. The most basic approach is file-based grouping, in which a single file (including its data blocks, inode, and indirect pointers) is treated as the logical unit of data; however, with this technique a user may find that some files in a directory are unavailable while others are not, which may cause frustration and confusion. Other groupings preserve more meaningful portions of the file system volume under failure. With directory-based grouping, D-GRAID ensures that the files of a directory are all placed within the same unit of fault containment. Less automated options are also possible, allowing users to specify arbitrary semantic groupings which D-GRAID then treats as a unit.
Load balance: With fault-isolated placement, instead of placing blocks of a file across many disks, the blocks are isolated within a single home site. Isolated placement improves availability but introduces the problem of load balancing, which has both space and time components.
In terms of space, the total utilized space in each disk should be maintained at roughly the same level, so that when a fraction of disks fail, roughly the same fraction of data becomes unavailable. Such balancing can be addressed in the foreground (i.e., when data is first allocated), the background (i.e., with migration), or both. Files (or directories) larger than the amount of free space in a single disk can be handled either with a potentially expensive reorganization or by reserving large extents of free space on a subset of drives. Files that are larger than a single disk must be split across disks.
More pressing are the performance problems introduced by fault-isolated data placement. Previous work indicates that striping of data across disks is better for performance even compared to sophisticated file placement algorithms [15,48]. Thus, D-GRAID makes additional copies of user data that are spread across the drives of the system, a process which we call access-driven diffusion. Whereas standard D-GRAID data placement is optimized for availability, access-driven diffusion increases performance for those files that are frequently accessed. Not surprisingly, access-driven diffusion introduces policy decisions into D-GRAID, including where to place replicas that are made for performance, which files to replicate, and when to create the replicas.
Meta-data replication level: The degree of meta-data replication within D-GRAID determines how resilient it is to excessive failures. Thus, a high degree of replication is desirable. Unfortunately, meta-data replication comes with costs, both in terms of space and time. For space overheads, the trade-offs are obvious: more replicas imply more resiliency. One difference between traditional RAID and D-GRAID is that the amount of space needed for replication of naming and system meta-data is dependent on usage, i.e., a volume with more directories induces a greater amount of overhead. For time overheads, a higher degree of replication implies lowered write performance for naming and system meta-data operations. However, others have observed that there is a lack of update activity at higher levels in the directory tree [35], and lazy update propagation can be employed to reduce costs [43].