ElasticLLM: An Adaptive and Cost-Aware Infrastructure Framework for Efficient Large Model Serving

The deployment of large language models (LLMs) in production environments faces significant challenges related to computational efficiency, resource utilization, and operational costs. Existing serving systems often struggle with dynamic workload patterns, resulting in either resource over-provisioning or performance degradation. We present ElasticLLM, an adaptive infrastructure framework that dynamically optimizes resource allocation for LLM serving through intelligent scheduling, cost-aware optimization, and adaptive batching strategies. Our system monitors real-time performance metrics and automatically adjusts computational resources to maintain service quality while minimizing costs. Experimental evaluations on multiple LLM architectures demonstrate that ElasticLLM achieves up to 1.48× improvement in throughput, 33% reduction in P95 latency, and 38% cost savings compared to static allocation approaches, while maintaining consistent service quality under varying workload conditions.