Maximum repair load
Maximum repair load is the largest per-node send or receive volume during a repair operation.
For each participating node, count incoming and outgoing payload separately and take the largest value across both directions and all nodes; do not add a node's send and receive volumes. The unit is bytes per repair, not bytes per second. [1]
This distinguishes a bottleneck's traffic from aggregate repair traffic. A lower total does not guarantee a smaller maximum node load or a shorter repair. Input reuse also matters: one sub-block sent to a destination can serve multiple calculations there. [1]
Hesela enumerates all 65,536 placements of eight computation vertices in one four-node repair graph. Source locations and the replacement collector are fixed. The downloadable calculation checks transfer conservation and deduplicates by source sub-block and destination. It measures neither decoding correctness nor elapsed time. [2]
Relations
- related
- Minimum-storage regenerating code, Subpacketization, Repair-read amplification, Repair locality