We estimate, by simulation, the cell-loss rate in an ATM switch modeled as a queueing network. Cell losses are rare events, so estimating their frequency by simulation is hard. We experiment with importance sampling as a mean of improving the simulation efficiency in that context.