libfabric

Overview

Libfabric is an open-source, low-level network API (part of the OpenFabrics Interfaces project) designed from the outset to be connectionless and to enable unordered operations. It runs over InfiniBand adapters, in HPE's Exascale supercomputer systems, and across multiple vendors' networks with radically different architectures. The Ultra Ethernet Consortium chose libfabric as the application-facing API for Ultra Ethernet, meaning AI and HPC application stacks (PyTorch, TensorFlow), collective communication libraries (CCLs, NCCL/RCCL), shared-memory libraries (SHMEM), and MPI frameworks require no modification to run over Ultra Ethernet. Keith D Underwood (HPE) was present when libfabric was created and confirms that although a connection-oriented path exists in the API, essentially nobody uses it.