The recent advances in downscaling and weather generators offers the opportunity to step away from the synthetic design hyetograph strategy for water infrastructure design under extreme events. A novel practice based on intensive sampling of extreme events, i.e., exploring the range of possible extreme rain with respect to their occurrence probability, was investigated. Such a practice allows for a more reliable picture of the behaviour of systems during extreme rainfall events for variable initial conditions. However, in order to apply such a method at a catchment scale where the computational cost is higher, there is still a need to investigate the number of hyetographs necessary to approach the distribution of performance. In this study, the sensitivity of the distribution-based performance indicators to the number of sampled events was investigated. Methods are also tested to reduce uncertainty linked to the number of events by improving the sampling strategy (based on events characteristics or surface answer metamodeling). In order to get a representative answer, the approaches are tested on 3 fast-simulation green roofs models with different behaviour: i) an extensive green roof, ii) a detention-based green roof, and iii) a discharge regulated green roof (subject to sudden failure once the limit is reached). The results show that the number of sampled events has to be chosen in accordance with i) the performance indicator nature, ii) the acceptable range of uncertainty, and iii) prior knowledge on the system behaviour.