RLQ: Workload Allocation With Reinforcement Learning in Distributed Queues

Distributed workload queues are nowadays widely used due to their significant advantages in terms of decoupling, resilience, and scaling. Task allocation worker nodes distributed queue systems is typically simplistic (e.g., Least Recently Used) or uses hand-crafted heuristics that require task-specific information task resource demands expected time execution). When such not available node capabilities homogeneous, the existing placement strategies may lead unnecessarily large execution timings usage costs. In this work, we formulate problem Markov Decision Process framework, which an agent assigns tasks resource, receives a numerical reward signal upon completion. Our adaptive learning-based solution, Reinforcement Learning based Queues ( xmlns:xlink="http://www.w3.org/1999/xlink">RLQ ), implemented integrated with popular Celery queuing system for Python. We compare against traditional solutions using both synthetic real traces. On average, workloads, reduces cost by approximately 70%, factor at least 3×, waiting almost 7×. Using traces, observe improvement about 20% cost, around 70% time, reduction 20× time. also strategy inspired E-PVM, state-of-the-art solution Google's Borg cluster manager, showing able outperform it five out six scenarios.