DEVELOPING AN EFFICIENT SCHEDULING MODEL FOR CLOUD-BASED AI TRAINING USING PREDICTIVE WORKLOAD PATTERNS

Abstract

Cloud-based AI training introduces unique challenges in managing computing resources efficiently under dynamic workloads. Predictive workload models can anticipate resource needs and optimize scheduling, reducing latency and energy consumption. This paper proposes a predictive scheduling model integrating LSTM-based workload forecasting and dynamic resource allocation tailored for AI training in cloud environments. We evaluate its performance against conventional models using real trace data, showing significant improvements in resource utilization and training throughput.