Optimal-access repair
Optimal-access repair attains a code's repair-bandwidth bound by reading and forwarding only the stored symbols needed from its helpers.
At the minimum-storage point, any k of n nodes recover the file. Symmetric single-node repair from d helpers, with k <= d < n, has a download bound of dL/(d-k+1) symbols for L symbols per node. Optimal access attains this with the same number of helper-symbol reads. State the failed-node class and eligible helper sets: systematic-node optimality alone does not cover parity failures. [1]
A six-node, four-data-node example stores two GF(7) symbols per node. Its published repair schedule sends five symbols for any one lost data node. Repairing N1 or N2 reads five stored symbols; repairing N3 or N4 reads ten to form five combinations. Both schedules meet the five-symbol bandwidth bound, but their access costs differ. These counts follow from the verified linear maps, not a disk benchmark. [3]
Forwarding selected symbols needs no helper-side arithmetic, but reconstructing the failed node can still require decoding. At the device boundary, alignment and fragmented sub-chunks can make physical reads larger than the logical symbol budget. [2]
Relations
- related
- Minimum-storage regenerating code, Repair-by-transfer, Subpacketization, CLAY code, Repair-read amplification, Maximum distance separable code
Limitations
- A symbol-access optimum does not establish the fastest repair: request layout and I/O granularity also affect cost. [2]
Citations
- Tamo, Wang and Bruck. Access vs. Bandwidth in Codes for Storage (2013), arXiv:1303.3668v1, sections I-II and V.
- Vajha et al. Clay Codes: Moulding MDS Codes to Yield an MSR Code. FAST 2018, sections 2 and 5.2.
- Hesela: exact GF(7) verification of the six-node repair example and its logical read counts.
Further reading
Sources & caveats
- The primary preprint's Figure 1 coefficients and repair schedule were reconstructed independently. No throughput, latency or physical-device measurements are inferred.
- Tests check every four-node subset for full rank and every data-node repair on a field basis; linearity extends correctness to all 7^8 input messages.
- The finite example covers the published systematic-node schedule only, not all possible repair algorithms, parity-node repair or arbitrary helper sets.