The Scotch software computes process-processor mappings by assigning recursively parts of the process graphs to parts of the target graphs. To date, while regular target architectures can be described using pre-coded routines, irregular architectures or parts of regular ones require O(P^2) data structures, which makes them unpractical for very big machines. We will present a new, multilevel description of target architectures that alleviates this problem, trading-off memory for run time.