<p>As cloud computing continues to become more popular, efficient task scheduling on virtual machines has become fundamental to optimizing resource use, lowering expenses, and minimizing delays. Traditional scheduling algorithms (FCFS, SJF) fail to adapt to dynamic situations and complex task dependencies. This paper presents LDRLLSTM, a lightweight deep reinforcement learning framework that integrates single-layer LSTM with Deep Q-Network (DQN) principles for adaptive cloud task scheduling. Differentiating from prior DRL schedulers which typically overlook concurrent multi-objective QoS constraints or require prohibitively deep architectures, LDRLLSTM is designed with three explicit design choices: (i) a single-layer LSTM encoder (32 units) enabling CPU-only training, (ii) a multi-objective reward function explicitly balancing latency (w<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_1\)</EquationSource> </InlineEquation>=0.5), throughput (w<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(_2\)</EquationSource> </InlineEquation>=0.3), and reliability (w<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(_3\)</EquationSource> </InlineEquation>=0.2) constraints with formal SLA thresholds, and (iii) rigorous validation of scalability across production-scale workloads (100–100,000 tasks). Empirical validation demonstrates LDRLLSTM achieves 98.0% task success rate (±0.6% CI), 85 ms mean response time (±3.8% CI), and 94% VM utilization on 10,000-task Google Cloud dataset–competitive with or superior to recent SOTA methods (PPO, A3C, SAC, Transformer) while maintaining 60-200<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> parameter efficiency. Critical innovation: the model trains on CPU in 18 min with 53% lower cost than GPU alternatives, enabling practical deployment in resource-constrained environments (private clouds, edge datacenters) where GPU infrastructure is unavailable or cost-prohibitive. Comprehensive evaluation includes: (1) statistical validation with 95% confidence intervals and ANOVA (<i>p</i><0.001), (2) scalability analysis from 100 to 100,000 tasks with graceful degradation, (3) comparison against 12 baseline methods including recent SOTA, (4) GPU vs. CPU cost-benefit analysis, and (5) interpretability analysis with feature importance and decision attribution. This work contributes a practical, cost-efficient QoS-aware scheduler suitable for both academic research and production cloud environments.</p>

QoS aware deep reinforcement learning technique for task scheduling in cloud computing

期刊论文
出版模式：

开放获取
发表日期：

2026年3月30日

查看全文

Farheen Naaz,
Ayes Chinmay,
Anmol Pattanaik

摘要

As cloud computing continues to become more popular, efficient task scheduling on virtual machines has become fundamental to optimizing resource use, lowering expenses, and minimizing delays. Traditional scheduling algorithms (FCFS, SJF) fail to adapt to dynamic situations and complex task dependencies. This paper presents LDRLLSTM, a lightweight deep reinforcement learning framework that integrates single-layer LSTM with Deep Q-Network (DQN) principles for adaptive cloud task scheduling. Differentiating from prior DRL schedulers which typically overlook concurrent multi-objective QoS constraints or require prohibitively deep architectures, LDRLLSTM is designed with three explicit design choices: (i) a single-layer LSTM encoder (32 units) enabling CPU-only training, (ii) a multi-objective reward function explicitly balancing latency (w \(_1\) =0.5), throughput (w \(_2\) =0.3), and reliability (w \(_3\) =0.2) constraints with formal SLA thresholds, and (iii) rigorous validation of scalability across production-scale workloads (100–100,000 tasks). Empirical validation demonstrates LDRLLSTM achieves 98.0% task success rate (±0.6% CI), 85 ms mean response time (±3.8% CI), and 94% VM utilization on 10,000-task Google Cloud dataset–competitive with or superior to recent SOTA methods (PPO, A3C, SAC, Transformer) while maintaining 60-200 \(\times \) parameter efficiency. Critical innovation: the model trains on CPU in 18 min with 53% lower cost than GPU alternatives, enabling practical deployment in resource-constrained environments (private clouds, edge datacenters) where GPU infrastructure is unavailable or cost-prohibitive. Comprehensive evaluation includes: (1) statistical validation with 95% confidence intervals and ANOVA (p<0.001), (2) scalability analysis from 100 to 100,000 tasks with graceful degradation, (3) comparison against 12 baseline methods including recent SOTA, (4) GPU vs. CPU cost-benefit analysis, and (5) interpretability analysis with feature importance and decision attribution. This work contributes a practical, cost-efficient QoS-aware scheduler suitable for both academic research and production cloud environments.