In the recently proposed distributed computing approach to protein folding a very large number of short independent simulations is performed. Using this method, folding events on a time scale orders of magnitude shorter than the experimental one have been reported. However, it has also been observed that the folding process is not an elementary kinetic step and that the presence of initial lag phases can bias short simulations toward atypical pathways. We study here a 20-residue three-stranded antiparallel beta-sheet peptide whose equilibrium properties can be characterized by atomistic molecular dynamics simulations. We found that the folding rate of this peptide is estimated correctly by the distributed computing approach when trajectories > approximately 1/100 of the equilibrium folding time are considered. We also found that the fastest folding events occur through high-energy pathways, which are unlikely under equilibrium conditions. These very fast folding pathways do not relax within the equilibrium denatured state that is stabilized by the transient presence of both native and non-native interactions, and they are characterized by the nearly simultaneous formation of the two beta-hairpins and a very small number of non-native contacts.