Online Learning of Weakly Coupled MDP Policies for Load Balancing and Auto Scaling

TL;DR

This paper develops a new model and algorithms for load balancing and auto scaling in systems with bursty traffic, using weakly coupled MDPs and a two-timescale LP-based approach for online learning and policy optimization.

Contribution

It introduces a novel weakly coupled MDP model for load balancing and auto scaling, along with a tractable relaxed LP formulation and an online learning algorithm.

Findings

Effective LP relaxation for complex control problems

Successful online policy learning in dynamic environments

Improved scalability for load balancing and auto scaling

Abstract

Load balancing and auto scaling are at the core of scalable, contemporary systems, addressing dynamic resource allocation and service rate adjustments in response to workload changes. This paper introduces a novel model and algorithms for tuning load balancers coupled with auto scalers, considering bursty traffic arriving at finite queues. We begin by presenting the problem as a weakly coupled Markov Decision Processes (MDP), solvable via a linear program (LP). However, as the number of control variables of such LP grows combinatorially, we introduce a more tractable relaxed LP formulation, and extend it to tackle the problem of online parameter learning and policy optimization using a two-timescale algorithm based on the LP Lagrangian.