With the current trend of multiprocessor machines towards more and more hierarchical architectures, exploiting the full computational power requires careful distribution of execution threads and data so as to limit expensive remote memory accesses. Existing multi-threaded libraries provide only limited facilities to let applications express distribution indications, so that programmers end up with explicitly distributing tasks according to the underlying architecture, which is difficult and not portable. In this article, we present: (1) a model for dynamically expressing the structure of the computation; (2) a scheduler interpreting this model so as to make judicious hierarchical distribution decisions; (3) an implementation within the Marcel user-level thread library. We experimented our proposal on a scientific application running on a ccNUMA Bull NovaScale with 16 Intel Itanium II processors; results show a 30% gain compared to a classical scheduler, and are similar to what a handmade scheduler achieves in a non-portable way.