In this paper we propose a new distributed hash table model called auto-adaptive distributed hash table (AA-DHT). This model uses a distributed profiling of the nodes of the DHT to dynamically adapt the size of the index tables in order to reduce both the message cost and the request latency. This work is an evolution of the architecture for a P2P computing model described by Dury (2004), We detail the auto-adaptive model, the protocols we implemented and tested and we give experimental results of the architecture in simulated networks of up to 640 nodes.