PRS: Predication-Based Replica Selection Algorithm for Key-Value Stores

Abstract

The tail latency of end-user requests, which directly impacts the user experience and the revenue, is highly related to its corresponding numerous accesses in key-value stores. The replica selection algorithm is crucial to cut the tail latency of these key-value accesses. Recently, the C3 algorithm, which creatively piggybacks the queue-size of waiting keys from replica servers for the replica selection at clients, is proposed in NSDI 2015. Although C3 improves the tail latency a lot, it suffers from the timeliness issue on the feedback information, which directly influences the replica selection. In this paper, we analysis the evaluation of queue-size of waiting keys of C3, and some findings of queue-size variation were made. It motivate us to propose the Prediction-Based Replica Selection (PRS) algorithm, which predicts the queue-size at replica servers under the poor timeliness condition, instead of utilizing the exponentially weighted moving average of the state piggybacked queue-size as in C3. Consequently, PRS can obtain more accurate queue-size at clients than C3, and thus outperforms C3 in terms of cutting the tail latency. Simulation results confirm the advantage of PRS over C3.