Metropolis sampling is based on a physical analogy, and it is able to generate samples from almost any arbitrary probability distribution by simulating a stationary Markov process. During 3D rendering, we have to sample the high-dimensional space of light paths connecting light sources to the virtual camera via scattering events on various surfaces, and we have to do it with importance sampling based on the energy being carried. The Metropolis method generates new samples by perturbing the current sample, then randomly accepting or rejecting the new sample. Perturbation does not affect the stationer distribution, which is guaranteed by setting the acceptance probability. However, the rate of convergence strongly depends on the perturbation strategy. This thesis presents the GPGPU based implementation of a global illumination algorithm employing Metropolis sampling built on the OptiX framework, and the performance evaluation of different perturbation strategies.