Parallel genetic algorithms on hybrid servers: Design, implementation, and optimization for performance and energy
Amr Abdelhafez, Ravi Reddy Manumachu, Alexey Lastovetsky · Swarm and Evolutionary Computation · 2025
Parallel Genetic Algorithms (PGAs) have been widely applied to accelerate solutions for real-world problems such as energy optimization in building constructions, data preprocessing and model selection steps in data mining, real-time control of multilevel inverters in electronics, land-use planning, nanoscience, optimal power flow in power systems, and road traffic management. The state-of-the-art research proposes PGAs optimized solely for performance and for solving optimization problems on a multicore CPU, GPU, or clusters of multicore CPUs. However, no research has analyzed PGAs for heterogeneous hybrid platforms comprising multicore CPUs and multiple accelerators that utilize all computing devices in parallel. Furthermore, no definitive comparative research comprehensively investigates the energy consumption of PGAs in hybrid systems versus multicore CPUs or GPUs. We address the above gaps in the prior art in this work. First, we present a novel parallelization approach (HPIGA) tailored for heterogeneous hybrid platforms, featuring a portable implementation that utilizes all available computational devices, including multicore CPUs and GPUs. We conduct a comprehensive investigation into the performance and energy profiles of this approach. We compare it with three other traditional parallel approaches across a range of dimensions, varying from 100 dimensions and up to 5000 dimensions. The results showed HPIGA’s competitive energy consumption behavior and promising performance compared to other traditional approaches under the study. Moreover, we formulate a bi-objective optimization problem of a PGA employing a parallel island model and executing on a hybrid server comprising p compute devices. The problem has two objectives: performance and energy. The decision variable used in our bi-objective optimization problem is workload distribution, which is proportional to the number of islands. We study the efficacy of our proposed PGA on a hybrid server platform with an Intel Icelake multicore CPU and two Nvidia A40 GPUs, analyzing execution time and dynamic energy profiles under two power governors. The resulting Pareto front graphs provide valuable insights, serving as crucial benchmarks for the future development and use of efficient, energy-aware optimization techniques across diverse computational devices.