Multi-Agent Negotiation for Adaptive Consumption of Continuous Allocation

In this paper, we address the problem of continuous allocation of concurrent jobs, composed of situated tasks, which underpins the distributed deployment of the MapReduce design pattern on a cluster. We propose a multi-agent strategy designed to minimise the mean flowtime of jobs through a composite agent architecture that enables negotiation and consumption to proceed concurrently. Unlike classical auction-based approaches, our method relies on bilateral negotiations between agents to reallocate tasks dynamically. This allows the multi- agent system to adapt on the fly to disruptive events, such as execution delays or job arrivals, without requiring an explicit model of the environment. Our experiments show that the proposed strategy is effective: (1) it significantly reduces rescheduling time, (2) improves mean flowtime, (3) does not penalise consumption, (4) is robust to execution hazards such as node slowdowns, and (5) adapts seamlessly to the release of new jobs. These results confirm the robustness and practical efficiency of negotiation-driven continuous allocation.