Optimization of Cloud Computing Resource Allocation Using Evolutionary Algorithms

Abstract

Cloud computing depends on efficient allocation of heterogeneous virtualized resources to workloads whose processing demands vary over time. Static or rule-based allocation can leave fast virtual machines underused while slower instances become overloaded, increasing application completion time and weakening quality-of-service performance. This methodology paper develops and evaluates an evolutionary resource-allocation framework in which task-to-virtual-machine mappings are optimized using a Genetic Algorithm (GA) and a discrete Particle Swarm Optimization (PSO) method. The objective combines normalized makespan, monetary execution cost, and load imbalance so that a scheduler does not optimize one measure at the expense of all others. A reproducible simulation is defined with 120 independent tasks and 10 heterogeneous virtual machines ranging from 1,200 to 5,200 MIPS. Evolutionary solutions are compared with Round-Robin and random allocation. Across 12 independent optimization runs, the GA obtains a mean makespan of 61.28 s and the PSO method 61.21 s, compared with 157.16 s for Round-Robin. Mean resource utilization increases from 48.1% under Round-Robin to about 98.3% under the evolutionary methods, while modeled execution cost changes only slightly. The results demonstrate why population-based search is suitable for the combinatorial allocation problem: it can explore many candidate mappings and rapidly converge toward balanced schedules on heterogeneous resources. The paper provides a clear experimental procedure that can be implemented directly in CloudSim or a comparable simulator for larger workloads and additional service-level constraints.