AGENT: An elitism-guided evolutionary framework for enhanced task allocation performance in heterogeneous cloud systems

Overview

This study addresses task allocation in Infrastructure as a Service cloud environments by proposing AGENT, a genetic algorithm framework that combines size-preserving elitism with nonlinear parameter adaptation for heterogeneous virtual machine scheduling. The research targets the NP-complete scheduling problem in cloud data centers with the objective of minimizing makespan, defined as the total execution time required to process all tasks across allocated resources.

Methods and approach

AGENT integrates three principal innovations: (1) size-preserving elitism that maintains population size while ensuring monotonic fitness improvement through selective retention of superior solutions, (2) feedback-based nonlinear parameter adaptation driven by explicit success and failure counters rather than population statistics or fitness proportionality, and (3) a multi-task-per-VM allocation model reflecting operational cloud elasticity. The framework was evaluated using CloudSim Plus simulation with Amazon EC2 virtual machine configurations to assess performance against baseline algorithms including Min-Min, Max-Min, AIGA, HAGA, and SGA across synthetic workload datasets of varying scales.

Key Findings

Experimental validation demonstrated makespan improvements ranging from 3.14% to 28.89% relative to baseline heuristics across tested datasets. The algorithm exhibited scalability with consistent near-optimal performance on workloads of different sizes. Secondary analysis indicates that reduced makespan corresponds to decreased VM active time, suggesting potential energy consumption improvements, though direct energy measurements were not conducted in this investigation.

Implications

The proposed elitism-guided approach demonstrates methodological advancement in adaptive genetic algorithm design for cloud task scheduling by decoupling parameter tuning from aggregate population statistics through explicit progress indicators. The multi-task-per-VM model provides closer alignment with operational cloud infrastructure characteristics compared to conventional single-task-per-VM abstractions, improving the relevance of scheduling solutions to real-world deployment scenarios. Future research directions include direct measurement and optimization of energy consumption metrics alongside makespan minimization, validation on heterogeneous workload distributions beyond synthetic benchmarks, and investigation of algorithm performance under dynamic resource availability conditions typical of production cloud environments.