Programming Model Extensions for Resilience
Research topic and goals
The reliability of high performance computing systems is predicted to worsen given current technology trends, where several errors per day are expected to occur in exascale systems. Consequently, the software stack that targets HPC systems needs to tolerate faults in order to cope with the instability of such systems. Particularly, current parallel programming models, which assume that applications will execute on failure-free systems, require resilience extensions. The most popular models used to program parallel systems are MPI for distributed memory and thread-based models, such as OpenMP, for shared-memory. Several works have studied faulttolerance methods for MPI applications at the process level such as the User-Level Fault Mitigation (ULFM) effort at ANL. Recently, the need for finer-grained levels of resilience arose in the community because of the space and time overheads of coarse-grained approaches. Consequently, research in this direction has become more active leading to new methods such as the rollbackrecovery method applied to asynchronous tasks in the OmpSs programming model. Despite recent advances, there exists no programming model that offers high degrees of resilience and low overheads for hybrid MPI+threads applications. This collaboration aims at providing resilience extensions for hybrid MPI+threads programming models at multiple levels. At the thread-level, errors are contained within the failing threads except if recovery is not possible, in which case propagating the failure to the MPI process will be necessary. The system then will rely on process-level fault-tolerance techniques to mitigate the failure.
Results for 2015/2016
There were interactions through telecon between the collaborators which involved discussions about ULFM for process-level resilience and the feasibility of thread-level fault-tolerance extensions.
Visits and meetings
There were frequent telecon meetings since the beginning of the collaboration. There are no planned visits yet.
Impact and publications
None yet.
Future plans
We plan to push further the investigation of fine-grained fault-tolerance support in hybrid MPI+threads programming models. The goal is to answer several open questions including, but not limited to, the degree of granularity suitable for the fine-grained level (threads or tasks), define the interaction between the coarse-grained and fine-grained levels of fault-tolerance, the way communication issued by a failed thread should be handled (flushed or ignored), and investigate the role and necessity of using robust synchronization mechanisms, such as robust mutexes, by the underlying multithreaded runtime system.