An Efficient Workload Prediction with Minimized SLA Violations in Cloud Computing using Deep Learning Model

Abstract

Accurate workload forecasting is crucial for optimizing resource distribution and minimizing cloud Service Level Agreement (SLA) breaches.  Due to their sequential and linear nature, existing models like LSTM, CNN, and their variants fail to learn complex temporal patterns and relationships, resulting in inferior prediction accuracy and SLA violations.  This paper presents a novel BiLSTM-based Multi-modal Attention Model (BiLSTM-MA) to address the challenges.  The hybrid model integrates BiLSTM networks with a multi-modal attention mechanism to enhance feature extraction from diverse cloud workload data.  Comprehensive trials demonstrate that BiLSTM-MA markedly outperforms leading models, including DLSTM, CNN-LSTM, mLSTM, and LSTM, across several parameters such as SLA violations, energy consumption, average CPU utilization, RMSE, and RMSPE.  The results indicate that the proposed model outperforms all baseline models across all workload.  The proposed model exhibits the lowest average CPU use, incurs the fewest SLA violations, and demonstrates minimal energy consumption, signifying its superior resource management capabilities while properly forecasting demand.  The mean CPU usage for BiLSTM-MA within the workload is 64%, indicating a reduction in RMSE error rates of 8.14%, 46%, 48%, and 55.39% compared to baseline approaches.