Skip to Main Content
Checkpointing is today's common mean for dealing with transient failures in supercomputers. However, the effectiveness of checkpointing and recovery protocols under the assumption that failures may happen during their operation is not well understood. We present an evaluation of the checkpointing and recovery based on Sender Based Message Logging protocols (SBML). We evaluate it by means of a model which is gathered from an extensive field data campaign performed on the SCOPE supercomputer at the University of Naples. A comprehensive model is built to evaluate reliability, scalability and performance of SBML. The proposed model takes into account failures during the checkpointing and recovery. Result provide insights on the limit of the number of nodes that can be allocated to the same job considering i) the overhead for the distributed coordination for rollback, and ii) network resources (e.g. bandwidth).