This work aims to present a computational structure of clusters of graphic calipers using a
recent distributed memory programming model, XcalableMP. Distributed memory parallel
programming often takes advantage of division of computing work among system CPUs, often
using message exchange mechanisms such as MPI, but since the discovery and growth of GPU
computing new possibilities have also emerged in the sense of To organize machines equipped
with GPUs in parallel computing environments, in order to obtain the main advantages of
both, mainly in the significant gain of computational performance. Thus, the work developed a
cluster architecture of graphic accelerators with the objective of obtaining computational gain
in the execution of genetic algorithms. The XcalableMP model was used as the distributed
memory process manager, with OpenACC as the GPU programming model forming the entire
framework of hybrid programming enablement. The genetic algorithms were executed and tested
by untangling the computational performance gain in execution in this structure when compared
to the sequential execution in CPU and execution in only one GPU.
|
