Keyun Cheng

Having Your Cake and Eating It Too: Jointly Optimal Erasure Codes for I/O, Storage, and Network-bandwidth

Download

FAST, 2015

Summary

Existing practical MSR codes fails to solve the high I/O overhead during repair. This paper introduces product-matrix-MSR codes that significantly reduces repair I/O, typically 5x, while retaining optimality with respect to storage/reliability/network bandwidth.

PM-RBT: Product-Matrix-Repair-By-Transfer

Details

  1. PM-MSR codes: a class of linear codes. PM-vanilla codes are not optimal with respect to the amount of data read during reconstruction. Reconstruction vector are mostly non-zero, not sparse, thus introduces high overhead.

  2. RBT-helper is introduced. RBT means if a node is used to transfer but without any computation. The methodology behind is: a lower bound on the network transfers is also a lower bound on the amount of data read. Separates RBT blocks and others. Use the others for the computation.

  3. An algorithm is presented for RBT helpers’s assignment. It’s doing:

Strength

Compared with RS,

  1. Both PM-vanilla nad PM-RBT, network bandwidth significantly goes lower.

  2. PM-RBT Significantly lowers I/O consumed and the data read during reconstruction of chunks.

  3. Similar decoding speed for PM-vanilla nad PM-RBT.

  4. Analysis of the RBT helper selection algorithm

Weakness

N/A