<p>As the volume of image, sound, and other data continues to grow exponentially, traditional CPUs are increasingly unable to handle the demands of high throughput. To address this challenge, GPU parallel computing is often employed to enhance performance. However, the efficient management of resources in CPU-GPU heterogeneous clusters remains a significant research challenge. Although current distributed resource management systems feature parallel schedulers that satisfy low-latency demands, they exhibit limitations in managing GPU resources and the inherent heterogeneity of CPU-GPU resources. Existing resource management systems often employ a coarse approach to GPU resource management and overlook the heterogeneity between CPUs and GPUs, resulting in lower utilization rates for both resource types. To address these challenges, this paper proposes a distributed resource management system that implements fine-grained management of CPU–GPU heterogeneous resources. The proposed system achieves fine-grained management of CPU–GPU heterogeneous resources through the integration of queue management technology and Cgroup resource limitation techniques, enabling efficient cluster resource sharing and utilization. The system is validated using the Spark computing framework, and the results demonstrate that the implemented strategies can effectively reduce application delay time resulting from resource conflicts, satisfying the design requirements and achieving the intended objectives. Additionally, a comparative analysis with the Mesos system is conducted to validate the rationality and effectiveness of the proposed system’s architecture design and key technologies.</p>