Core Concepts
The authors develop and analyze adaptive load balancing policies that can maintain a balanced distribution of tasks across server pools, even when the arrival rate of tasks is time-varying or the service time distribution is non-exponential. The policies integrate a threshold-based dispatching rule with online learning schemes to dynamically adjust the threshold in response to changes in the unknown offered load.
Abstract
The paper considers a service system with n identical server pools, where tasks with exponentially or Coxian distributed service times arrive as a (possibly time-inhomogeneous) Poisson process. To maintain a balanced distribution of the load across the server pools, the authors propose a threshold-based dispatching rule that assigns incoming tasks to server pools based on their current occupancy levels.
The key aspects of the paper are:
Dispatching rule:
Tasks are dispatched to server pools with fewer than a threshold number of tasks, chosen uniformly at random.
The threshold is dynamically adjusted over time to track the unknown offered load.
Learning schemes:
A basic scheme tracks the total number of tasks in the system to estimate the offered load.
A refined scheme uses more detailed occupancy information to adjust the threshold.
Analysis:
For the refined scheme with time-varying exponential service times, the authors prove that the threshold reaches an equilibrium value during intervals where the offered load is suitably bounded. This results in an asymptotically balanced distribution of the load.
For the basic scheme with time-homogeneous Coxian service times, the authors establish similar results on the threshold convergence and load balancing.
Methodology:
The authors develop a novel proof technique that leverages process-level dynamical properties through strong approximations, rather than relying on traditional fluid limit analysis.
This approach allows the authors to handle rapid threshold adjustments triggered by small occupancy fluctuations, and to characterize the asymptotic behavior without detailed state descriptors.
Extensions:
The results can be extended to other combinations of learning schemes and traffic scenarios, as discussed in the paper.
Overall, the paper presents adaptive load balancing policies that can effectively maintain a balanced distribution of tasks in large-scale systems with uncertain and time-varying demand, while requiring limited state information and communication overhead.