The migration of beryllium was studied in JET firstly to provide insight into the underlying transport processes and secondly to establish a reference for comparison to the upcoming bulk Be wall experiments (ILW). The JET wall was prepared in a state far from equilibrium by means of Be evaporation followed by a series of identical plasma discharges in which the relaxation of the wall composition towards the steady state situation was studied by measurement of wall erosion sources using visual range spectroscopy. The Be erosion flux at the midplane main chamber wall initially increased by a factor of up to 20 whereas the corresponding C flux decreased initially only by 50% with respect to the value before Be evaporation. From this observation one can infer that even extended Be evaporation does not lead to a uniform and closed Be layer on top of the carbon wall elements. The subsequent erosion of the Be layer, uncovering of substrate C and mixing with redeposited C occurs on a timescale of 120s in experiments where standard limiter phase start-up scenarios were used. After this time the Be flux in the limiter phase has already decreased to about a factor 5 above the reference value, whereas the C erosion flux has increased moderately to about 60% of the reference value. The corresponding decay of the Be flux in L-mode plasmas with early X-point formation and correspondingly short limiter phase is slower with a characteristic time constant of 400s. To quantitatively interpret the spectroscopic data, which represent only material gross erosion rates and their change over time due to the evolution of respective material fractions at the wall surface, a new integrated model has been developed, which allows describing the evolution of plasma exposed wall surfaces by impurity transport processes in fusion devices with different first wall materials. The experimentally observed time scale of the Be source evolution can be well reproduced by the model. Parameter studies reveal that the timescale is closely linked to the fraction of local redeposition vs. long range migration of eroded material. The simulations show also that computational grids for transport modelling, which do not extend flush to the first wall, will introduce significant errors in the predicted migration pattern.