Next generation HPC systems will be large scale distributed systems spread over wide area networks. Overlays are used in those systems to provide scalable and fault tolerant communication mechanisms. In such a context, providing highly available services to users is challenging. In this paper, we present Semias, a framework that provides stateful services with high availability and self-healing. Based on active replication on top of a structured overlay, Semias requires very few modifications of existing services. Semias self-healing mechanisms are designed to minimize the number of reconfigurations of replicated services while ensuring high availability. We have used Semias to make Vigne grid middleware services highly available. Experiments run on the Grid'5000 testbed show the performance and self-healing properties of the framework.