DeepSeek-Enabled Adaptive Heterogeneous Computing Platform: Architecture, Intelligent Scheduling, and Operation and Maintenance Optimization

Abstract

Traditional high-performance computing (HPC) platforms have significant deficiencies in heterogeneous resource integration, coping with the diversity of user requirements, and achieving real-time monitoring and optimization. To address these challenges, this paper designs and implements a heterogeneous resource fusion intelligent computing management platform based on DeepSeek. The platform uniformly manages heterogeneous computing powers such as CPUs, GPUs, and FPGAs through dynamic resource pooling technology, and combines intelligent scheduling algorithms driven by reinforcement learning and containerization technology (Slurm + Singularity) to achieve the co-scheduling of traditional HPC tasks and AI jobs. Particularly crucial is that the platform deeply integrates the DeepSeek-R1 large model to build a domain knowledge base, further supporting natural language interaction, script automatic generation, and fault diagnosis. Experimental results show that compared with traditional platforms, the utilization rate of heterogeneous resources increases by , the task response time is shortened by , and the efficiency of automated fault recovery is increased by 60%; the knowledge base achieves a problem coverage rate of 98.4% and a script generation success rate of 89.3%, and the user learning cost is reduced by more than 50%. This research provides a new paradigm for resource management in the "Ultra-Smart Fusion" era and will be extended to the cloud-edge-end collaborative architecture and the adaptation direction of memory-computation integrated hardware in the future.