Abstract--Clusters featuring the InfiniBand interconnect are continuing to scale. As an example, the "Ranger" system at the Texas Advanced Computing Center (TACC) includes over 60,000 cores with nearly 4,000 InfiniBand ports. The latest Top500 list shows 30% of systems and over 50% of the top 100 are now using InfiniBand as the compute node interconnect. As these systems continue to scale, the Mean-Time-Between-Failure (MTBF) is reducing and additional resiliency must be provided to the important components of HPC systems, including the MPI library. In this paper we present a design that leverages the reliability semantics of InfiniBand, but provides a higherlevel of resiliency. We are able to avoid aborting jobs in the case of network failures as well as failures on the endpoints in the InfiniBand Host Channel Adapters (HCA). We propose reliability designs for rendezvous designs using both Remote DMA (RDMA) read and write operations. We implement a prototype of our design an...
Matthew J. Koop, Pavel Shamis, Ishai Rabinovitz, D