Title: Optimal Data Placement for Stripe Merging in Locally Repairable Codes.
Conference (INFOCOM’22): Link
Journal (): Link
This paper studies redundancy transition for LRC, especially, merging multiple small LRC stripes to a large stripe. It first shows that random placement introduces high transition bandwidth, and propose two specific placement for two specific merge regimes, named aggregated merging and dispersed merging. They are provenly minimizing the cross-cluster scaling bandwidth, while preserving the repair bandwidth.
Formally prove that the random placement introduces high transition bandwidth.
Introduce two specific data placements for aggregated merging and dispersed merging, to minimize cross-cluster scaling bandwidth.
aggregated merging: fix (1) the number of global parities; (2) group size, merge the data and local parities
dispersed merging: fix (1) the storage redundancy and (2) group size, which is the initial n - k, and final n - k
Challenges: the merging bandwidth for two types of merging is high; specifically, the relocation bandwidth is high for aggregated merging; the recalculation bandwidth is high for dispersed merging, as it needs to calculate more parities.
Proof: 5 propositions to prove the issues for the two problems
Design:
Key idea: control the number of blocks to be placed in the same cluster, or the number of clusters for a stripe and placement.
Algorithms and tradeoff analysis: recalculation cost and migration cost for aggregated and dispersed merging represents the two extremes.
Optimality: any random placement can be converted into prescribed placement.
The parameter spaces supported is larger than INFOCOM’20. The two prescribed placements are achievable, which means they can converted from any placements.
Extensive experiments
N/A for now