In this paper we describe an mpiJava extension that implements a parallel checkpointing/recovery service. This checkpointing/recovery facility is transparent to applications, i.e. no instrumentation is needed. We use a distributed approach for taking the checkpoints, which means that the processes take their local checkpoints independently. This approach reduces communication between processes and there is not need for a central server for checkpoint storage. We present some experiments which suggest that the benefits of this extended MPI functionality do not have a significant performance penalty as a side effect, apart from the well-known penalties related to the local checkpoint generation. © Springer-Verlag Berlin Heidelberg 2006.