Differential Evolution for VNE-5G Scenarios

Vertical Parallelization of Differential Evolution Heuristic for Network Slicing in 5G Scenarios

“…In our earlier work [Gomes et al 2021] we set out some reasons for choosing the Genetic Algorithm (GA) as a good competitor: (i) it allows searching for the solution in a huge solution space; (ii) it does not restrict the search process to local domains; and (iii) it is flexible enough to be adapted to different scenarios. Moreover, the fitness function and chromosomal structure focus on adapting the heuristic GA to solve different problems.…”

“…A comprehensive modeling is presented in [Gomes et al 2021], which is the same we followed in this work. A network infrastructure is represented by a non-directed graph denoted by G = (V, E, β), in which G represents a physical network, V is a set of vertices in which v i ∈ V and represents a real device, an edge e i ∈ E, where E is a set of edges and it represents a real link.…”

“…Regarding the aspects mentioned above, in this work, we extend our previous work [Gomes et al 2021] in which it was the first to design and adapt the Differential Evolution (DE) to the VNE problem. We redesigned it to seek a way to reduce the runtime using a parallelization method, a method in which there is a lack in the literature regarding the evolutionary approaches.…”

“…Moreover, we magnified the embedding complexity considering node capacity, link bandwidth, and quality of services parameters such as delay and reliability. Further, the approach proposed in [Gomes et al 2021] has a small number of requisitions and does not seek to increase the acceptance rate.…”

“…Our previous work [Gomes et al 2021] reveals that the DE takes longer to run for the same number of repetitions and population size than the Genetic Algorithm (GA). Hence, in this paper, we enhanced these points: (i) we seek to improve the number of mapped requisitions; thus, we conceived a new fitness function; (ii) we endeavor to reduce the runtime of DE by parallelization; (iii) we adopt two more complex network topologies in the tests; (iv) we amplify the number of approaches from one to ten, they derive from Greedy, Stress, GA, and DE approaches; (v) the parallelization is not a silver-bullet, so we present in what conditions the parallelization of DE has advantage and disadvantage; and (vi) we propose two ways of initializing the first populations.…”