DPDNS Keynote

Summary form only given. In this talk, we will explore some recent results concern ing the execution of MPI applications on unstable environments. We will show that by extracting the fundamental characteristics of HPC application, we can design new fault tolerance approaches surpassing existing approaches. In particular, we will present a characterization of HPC applications and the design of a new family of fault tolerance protocols mixing the benefit of coordinated checkpointing and message logging protocols.