The data grid provides geographically distributed resources for large-scale applications. It generates a large set of data. The replication of this data in several sites of the grid is an effective solution for achieving good performance. In this paper we propose an approach of dynamic replication in a hierarchical grid that takes into account crash failures in the system. The replication decision is taken based on two parameters: the availability and popularity of the data. The administrator requires a minimum rate of availability for each piece of data according to its access history in previous periods, but this availability may increase if the demand is high on this data. We also proposed a strategy to keep the desired availability respected even in case of a failure or rarity (no-popularity) of the data. The simulation results show the effectiveness of our replication strategy in terms of response time, the unavailability of requests, and availability.