Keyun Cheng

Reading Notes: SRDS’22 XHR-Code

Title: XHR-Code: An Efficient Wide Stripe Erasure Code to Reduce Cross-Rack Overhead in Cloud Storage Systems

Conference (SRDS’22): Link

Journal (): Link

Summary

This paper proposes XHR-Code to address the multiple failures repair overhead from ECWide, which applies Azure-LRC under the wide-stripe hierarchical settings. It adds an intermediate layer of locality encoded by Hitchhiker between the XOR local parities and global parities. It also controls the storage overhead by enlarging the local group size, while keeping the number of racks small. Experiments shows its effectiveness.

Main Contributions

Details

Strength

• The Hitchhiker local parity encodes the data blocks horizontally from the rack level, which is new.

Weakness

• Overall, this paper has limited design novelty, which starts from ECWide, and adds a layer of local parity blocks similar to pyramid codes with three levels of hierarchy

• The design of using Hitchhiker to help multiple failure is definitely wrong
  • Hitchhiker by design ONLY helps single failure repair. This is absolutely wrong

• The design has many typos and makes the reading very difficult to understand. Also, the design is not accurately described especially for the Hitchhiker layer. What should be the number of rack groups? How many racks should one rack group contain? From the figures, it looks that every rack group has exactly z - 1 racks; however, the texts are not described clearly. Thus, the exact configurations appeared in the evaluation cannot be fully understood.

• For XHR-Code itself, how it makes the tradeoff between storage overhead, single parity block repair bandwidth, and multiple parity block repair bandwidth is not clear. Given the same storage overhead, if we add additional local parity blocks to protect the rack level fault tolerance, the number of racks is reduced, however, the local group size also becomes larger. Thus, the single parity block repair performance will also be degraded. This is not discussed in the paper.