Implementing Reliable Data Structures for MPI Services in High Component Count Systems
- Cite this paper as:
- Wozniak J.M., Jacobs B., Latham R., Lang S., Son S.W., Ross R. (2009) Implementing Reliable Data Structures for MPI Services in High Component Count Systems. In: Ropo M., Westerholm J., Dongarra J. (eds) Recent Advances in Parallel Virtual Machine and Message Passing Interface. EuroPVM/MPI 2009. Lecture Notes in Computer Science, vol 5759. Springer, Berlin, Heidelberg
High performance computing systems continue to grow: currently deployed systems exceed 160,000 cores and systems exceeding 1,000,000 cores are planned. Without significant improvements in component reliability, partial system failure modes could become an unacceptably regular occurrence, limiting the usability of advanced computing infrastructures. In this work, we intend to ease the development of survivable systems and applications through the implementation of a reliable key/value data store based on a distributed hash table (DHT). Borrowing from techniques developed for unreliable wide-area systems, we implemented a distributed data service built with MPI  that enables user data structures to survive partial system failure. The service is based on a new implementation of the Kademlia  distributed hash table.