Legio: fault resiliency for embarrassingly parallel MPI applications

Due to the increasing size of HPC machines, fault presence is becoming an eventuality that applications must face. Natively, MPI provides no support for execution past detection a fault, and this more constraining. With introduction ULFM (User Level Fault Mitigation library), it has been provided with possible way overtake during application at cost code modifications. intrusive in requires also deep understanding its recovery procedures. In paper we propose Legio, framework lowers complexity introducing resiliency embarrassingly parallel application. By hiding behind calls, library capable expose features transparent manner thus removing any integration effort. Upon failed nodes are discarded continues only non-failed ones. A hierarchical implementation solution proposed reduce overhead repair process when scaling towards large number nodes. We evaluated our solutions on Marconi100 cluster CINECA, showing introduced by negligible does not limit scalability properties MPI. Moreover, integrated real-world further prove robustness injecting faults.