We performed an on-sky MCAO experiment using 4 deformable mirrors (DMs) to analyze the relevance of their sequence to the residual wavefront error. Two DMs were conjugate to 4 and 8 km. The other two DMs were placed in pupil images upstream and downstream of the 4-km and 8-km mirrors. At any time, both high altitude DMs were active but only one pupil DM was active while the other one stayed flat. Firstly, we found that the MCAO control loops using either pupil DM were stable and robust. Dynamic misregistration, which was present for the first pupil DM, was not an immediate problem for the controller. We did not notice an apparent difference when repeatedly switching between the pupil DMs during the operation. A closer analysis of the contrast in the corrected images and AO telemetry indicates an advantage when the pupil correction was applied with the DM that was downstream of the high-altitude DMs. This finding is consistent in several data recorded at different days. The difference, however, is small. A more detailed analysis is probably needed to rule out error sources potentially not considered herein to draw a final conclusion on the optimal sequence of DMs in MCAO and its practical relevance.