Title: Practical Design Considerations for Wide Locally Recoverable Codes (LRCs)
Conference (FAST’23): Link
Journal (): Link
This paper compares different LRC constructions from a practical perspective, including the average data repair cost, average repair cost (including the global parities, which is more friendly for full-LRCs that can support local repair for global parities), and MTTDL. It proposes a construction called Uniform Cauchy LRC (which is very close to ATC’18) which can repair more failure patterns than the prior constructions in multiple failures where the number of failures is larger than the code distance. The comparisons are based on theoretical analysis and Monte-Carlo simulation experiments.
Compares state-of-the-art LRC constructions for wide stripes, including Azure-LRC, Azure-LRC+1, Optimal Cauchy LRC (Xorbas)
Reliability analysis using observed failures (via Monte Carlo simulations)
Analyze the theoretical properties of Optimal distance LRCs and MR-LRCs with maximal recovery property (not necessarily distance optimal)
Propose Uniform Cauchy LRC, a sub-optimal distance LRC that can tolerate more failure patterns with the number of failures larger than the distance, with a tradeoff of smaller code distance than the Optimal Cauchy LRCs.
Optimal-LRC: very close to ATC’18
Uniform Cauchy LRC: the coefficients of global parities are the same as Optimal-LRC; the coefficients of local parities changed.
One useful information: analyzing the multiple failure patterns in practice, which suggests that multiple failures appear more common than shorter codes. Supporting the multiple failures (especially on the system performance) for wide-LRCs needs more research.
The construction is almost the same as Optimal-LRC proposed in ATC’18, which makes the novelty very limited. The only difference is that the new construction can repair more failure patterns for multiple failures beyond the code distance. By the way, it makes the code distance lower than the original Optimal-LRC (ATC’18). Overall, I think the theoretical construction is very limited.
It only contains theoretical analysis. No real-world experiments are conducted.