Keyun Cheng

Reading Notes: IPDPS’23 multiple repair

Title: Boosting Multi-Block Repair in Cloud Storage Systems with Wide-Stripe Erasure Coding

Conference (IPDPS’23): Link

Journal (): Link

Summary

This paper combines the idea of centralized repair (collect data for repair at a specific node) and independent repair (repair pipelining) to solve the multiple repair problem in a heterogeneous network setting, under the wide-stripe setting. Evaluation under the heterogeneous network settings demonstrates the effectiveness of the combination.

Main Contributions

• Propose a theoretical model to combine the centralized repair and independent repair, such their repair time are equal.

• Extend to hierarchical settings and multiple node repair problem.

• System implementation based on OpenEC

Details

• Modeling for multiple block repair with RS codes
  • Centralized repair downloads k blocks to a node and distribute the repaired blocks to the new nodes
  • Independent repair uploads and downloads x blocks at each node and repair in parallel (same as repair pipelining)
  • A block can be physically divided into two parts, where for each part we adopt one repair approach (either centralized or independent)
  • Thus, the modeling can be used to calculate the tradeoff point, such that the time for centralized repair and independent repair are the same
    • By solving a linear equation
  • It considers the heterogeneous network setting, where the bandwidth between nodes are different (assume they are known in advance)
  • Based on the modeling (with given coding parameters, network bandwidth (upload and download) for each node, block size, etc.), we find the point where both of these two approaches takes the same time to finish
• Extension to hierarchical setting
  • Only focus on cross-rack network bandwidth
• Multi-node repair
  • Assign the repair on different nodes via a combination of LFS (frequency) or LRS (recently) approach
• Evaluation
  • Generates the heterogeneous network settings
  • Settings
    • Over EC2 with 88 nodes, 10Gbps network, 2vCPU, 8GiB memory

Strength

• Combines both approaches

Weakness

• Design (it’s good enough for a rank B paper)
  • Trivially combines both existing two approaches
  • Wide-stripe is not the focus
  • I think multiple repair is also not the focus. The independent repair approach for multiple repair has f times bandwidth consumption. The bandwidth it generates is much higher than the centralized repair
    • I think one experiment regarding the proportion of centralized repair and independent repair is strongly needed. Otherwise, it’s not clear how much bandwidth is trade-offed for the repair performance
• Evaluation
  • To me it has many problems
  • The network setting is called a dataset, but I would consider it as an list of upload and download bandwidth. I can’t find the dataset on the code repo
  • The redundancy is high (e.g., (64, 25)), I would not call it wide-stripes
  • Figure 8 has many typos
  • For exp6, the time breakdown is not reasonable
    • The operations performed are in parallel, thus we cannot directly count the time by separating the two sub-blocks