Implementation:
Making D-GRAID

We now discuss the prototype implementation of D-GRAID known as Alexander. Alexander uses fault-isolated data placement and selective meta-data replication to provide graceful degradation under failure, and employs access-driven diffusion to correct the performance problems introduced by availability-oriented layout. Currently, Alexander replicates namespace and system meta-data to an administrator-controlled value (e.g., 4 or 8), and stores user data in either a RAID-0 or RAID-1 manner; we refer to those systems as D-GRAID Levels 0 and 1, respectively. We are currently pursuing a D-GRAID Level 5 implementation, which uses log-structuring [39] to avoid the small-write problem that is exacerbated by fault-isolated data placement.

In this section, we present the implementation of graceful degradation and live-block recovery, with most of the complexity (and hence discussion) centered around graceful degradation. For simplicity of exposition, we focus on the construction of Alexander underneath the Linux ext2 file system. At the end of the section, we discuss differences in our implementation underneath VFAT.