Local Rendezvous Hashing for Balanced Loads and Minimal Churn
Analysis
This paper introduces Local Rendezvous Hashing (LRH) as a novel approach to consistent hashing, addressing the limitations of existing ring-based schemes. It focuses on improving load balancing and minimizing churn in distributed systems. The key innovation is restricting the Highest Random Weight (HRW) selection to a cache-local window, which allows for efficient key lookups and reduces the impact of node failures. The paper's significance lies in its potential to improve the performance and stability of distributed systems by providing a more efficient and robust consistent hashing algorithm.
Key Takeaways
“LRH reduces Max/Avg load from 1.2785 to 1.0947 and achieves 60.05 Mkeys/s, about 6.8x faster than multi-probe consistent hashing with 8 probes (8.80 Mkeys/s) while approaching its balance (Max/Avg 1.0697).”