Efficient GPU Memory Resource Scheduling Algorithm for Vehicle Detection Tasks in High Concurrent Scenarios

In response to the fierce competition and low utilization of GPU memory resources in high concurrency vehicle detection tasks at electric vehicle charging stations, this paper proposes a dynamic memory scheduling algorithm. This method integrates multiple GPU resources through GPU memory pooling technology, combined with improved greedy algorithm to achieve dynamic allocation of GPU memory and fragment optimization. The algorithm adopts a hierarchical architecture, monitors the status of GPU memory in real time, prioritizes scheduling low occupancy tasks and reserving emergency resources, and triggers the service replanning strategy with the minimum migration cost when GPU memory is insufficient, balancing resource utilization and service quality. Experiments have shown that in simulated charging station scenarios, compared to traditional static allocation strategies, this algorithm improves GPU memory utilization to 84%, achieves a task scheduling success rate of up to 89%, controls response delay within 54ms. This scheme effectively solves the problem of task queuing and resource contention in peak hours, provides an efficient and economical general framework for the deployment of high concurrency AI tasks in the edge computing scenario, and can be extended to intelligent transportation, industrial detection and other fields, helping to green and intelligent upgrade of intelligent infrastructure.