Beyond the Ring: Quantized Heterogeneous Consistent Hashing

Yoav Levi, Isaac Keslassy

Abstract: Consistent hashing (CH) is a crucial building block for load-balancers. It enables packets of the same flow to keep being mapped to the same server whenever possible. Load- balancers implement heterogeneous CH to deal with the varying server speeds. However, they mostly rely on the old Ring algorithm, suffering from practical scalability and stability issues, as well as from a lack of theoretical stability guarantees. This paper presents a new framework for heterogeneous CH. The framework relies on quantization using virtual servers, and on a min-max fairness-based mapping algorithm denoted M3. The paper establishes necessary and sufficient conditions to guarantee stability for any arbitrary heterogeneous server service rate. We also explain why M3 presents better scalability properties than all heterogeneous CH alternatives, including a faster key lookup rate and a lower memory footprint. Finally, evaluations show that M3 also offers a significantly increased stability region.