GRID computing is a more and more attractive approach. Its aim is to gather and to share the resources of a network like the
content, the storage or CPU cycles. A computational distributed system like SETI@home produces a power up to 70 TFlops whereas the current best parallel supercomputer BlueGene produces a power of 140 TFlops. Such a supercomputer costs very much contrary to a system like SETI@home. But the use of many computers to increase the global computational power involves several communication problems. We must
maintain the GRID communication in order to make any type of computation even though the network is volatile.
In this paper, we present a model to represent GRID applications and networks in order to show faults impacts. We present
a fully distributed solution based on a random walk to manage the topology of the GRID. No virtual structure needs to be maintained
and this solution works on asynchronous networks. We also present some simulations of our solution.