We first examine the space overheads due to selective meta-data replication that are typical with D-GRAID-style redundancy. We calculate the cost of selective meta-data replication as a percentage overhead, measured across all volumes of the HP trace data. We calculate the highest selective meta-data replication overhead percentage possible by assuming no replication of user data; if user data is mirrored, the overheads are cut in half.
| Level of Replication | |||
| 1-way | 4-way | 16-way | |
| ext21KB | 0.15% | 0.60% | 2.41% |
| ext24KB | 0.43% | 1.71% | 6.84% |
| VFAT1KB | 0.52% | 2.07% | 8.29% |
| VFAT4KB | 0.50% | 2.01% | 8.03% |
Table 1 shows that selective meta-data replication induces only a mild space overhead even under high levels of meta-data redundancy for both the Linux ext2 and VFAT file systems. Even with 16-way redundancy of meta-data, only a space overhead of 8% is incurred in the worst case (VFAT with 1 KB blocks). With increasing block size, while ext2 uses more space (due to internal fragmentation with larger directory blocks), the overheads actually decrease with VFAT. This phenomenon is due to the structure of VFAT; for a fixed-sized file system, as block size grows, the file allocation table itself shrinks, although the blocks that contain directory data grow.