A Dynamic Replica Quantity Adjustment Strategy Based on File Popularity Prediction

To address the inefficiency of static replica strategies in handling dynamically changing access heat in storage systems, this paper proposes a dynamic replica adjustment strategy based on file heat prediction for dynamic resource management. First, a multi-period weighted heat calculation method is defined, integrating short-, medium-, and long-term access patterns with file modification frequency as an auxiliary metric. Second, a CNN-LSTM-Attention model is constructed: CNN extracts local spatial-temporal features, LSTM captures long-term dependencies, and the attention mechanism dynamically weights key time steps for accurate heat prediction. Experimental results show that this model outperforms SVR, ARIMA, and single LSTM in prediction error. Based on this, a Dynamic File Adjustment (DFA) algorithm is proposed, combining a composite threshold trigger with priority-based resource allocation to adjust replica counts according to file heat levels. Compared with Hadoop’s default strategy, DFA improves both average response time and large-job completion time, demonstrating enhanced storage resource utilization and service quality.