Keyun Cheng

Reading Notes: ICPP’22 Repair placement for Optimal-LRC

Title: Toward Optimal Repair and Load Balance in Locally Repairable Codes

Conference (ICPP’23): Link

Journal (): Link

Summary

• For optimal-LRC in clustered storage systems, propose (1) a data partition scheme and (2) a node selection scheme to reduce cross-rack repair bandwidth and load balance.

Main Contributions

Details

• Scenario
  • Clustered storage
  • Azure-LRC, k is divisible by # of local groups
• Problem
  • data partitioning affects the cross-cluster network traffic
  • node selection can affect load balance
• Challenge
  • random data placement causes high cross-rack repair bandwidth
    • Specifically, it focuses on repairing the global parity block problem for ECWide
• Motivation: Minimize # of clusters for each group and blocks
  • Data and global parity blocks span less clusters, the cross-rack BW for repairing global parity block can be reduced (trivial)
  • Under careful partitioning, the cross-rack BW can be reduced to optimal for single and global repair
  • Random node and cluster selection incur storage and network load imbalance
    • Considering the access frequency: local parity blocks are more commonly accessed
    • Put hot data in less loaded nodes to improve load balance
    • Considering the heterogeneity of clusters and nodes (new nodes and new clusters are with higher storage capacity)
• Note: the motivating example under Motivation 1 is wrong
  • Specifically, the global parity cross-rack repair bandwidth is not two, but three. Three means that G1 and G2 does not help at all. There are two sets of linear independent equations, we cannot only decode for one.
• Design
  • Metrics
    • Storage load and bias
    • Network load and bias
  • Physical meaning: to measure the load imbalance (I think the description is not clear)
  • Partitioning
    • DRC and NRC are both minimized under single cluster fault tolerance
  • Node selection: select proper nodes in clusters for load balance
    • Measure the potential load of each partition and for each block

Strength

• Propose a rack and node placement scheme for Azure-LRC for reducing the cross-rack bandwidth and load balancing

Weakness

• In numerical analysis, there is no hotness related analysis for design 2 (node selection). The design idea is too high level and needs further elaboration

• The baseline is very trivial. Needs comparative study versus existing placement schemes, for example, ECWide.