Predictions in Multiserver and Networked Queueing Systems

Investigate how to incorporate predicted job sizes into scheduling and load balancing for multiserver and networked queueing systems, and develop theoretical analyses or guarantees that extend beyond empirical studies.

Background

Most existing work on using predictions in multiserver systems is empirical, and analysis is challenging even without predictions. The paper notes limited theoretical results, with one recent asymptotic throughput optimality result for SPJF with noisy estimates in M/GI/s+GI systems.

The authors emphasize that multiserver systems and networks present natural settings where predictions could be useful (e.g., load balancing with the power of d choices), but theoretical models and fluid-limit analyses that capture prediction use remain largely undeveloped.