The simulation of CMS raw data requires the random selection of one hundred and fifty pileup events from a very large set of files, to be superimposed in memory to the signal event. The use of ROOT I/O for that purpose is quite unusual: the events are not read sequentially but pseudo-randomly, they are not processed one by one in memory but by bunches, and they do not contain orthodox ROOT objects but many foreign objects and templates. In this context, we have compared the performance of ROOT containers versus the STL vectors, and the use of trees versus a direct storage of containers. The strategy with best performances is by far the one using clones within trees, but it stays hard to tune and very dependant on the exact use-case. The use of STL vectors could bring more easily similar performances in a future ROOT release.