The OpenFabrics Alliance, the OpenFabrics Interfaces Working Group, and the libfabric community are pleased to share that libfabric 2.6.0 is now available.
libfabric 2.6.0 continues the project’s steady evolution to support next-generation AI and HPC infrastructure. The release focuses on improving GPU networking, optimizing provider performance, expanding hardware support, and strengthening validation through enhanced testing. These updates help ensure libfabric remains a high-performance, portable communication layer for modern accelerator- and RDMA-based systems.
Key Updates in Version 2.6.0
The 2.6.0 release includes various fixes and improvements to the core and providers, including:
- shm provider
- Address some performance regressions introduced in v2.5.x
- verbs provider
- Add support for manual & auto GPU-NIC affinity setting
- efa provider
- Hardware counter support
- opx provider
- Improvement and fixes in dmabuf handling
- Fabtests
- Expand dmabuf-rdma test with multi-node runs and all-to-all communication pattern
About libfabric
Libfabric, also known as Open Fabrics Interfaces (OFI), is a framework focused on exporting high-performance networking services to applications. It specifically targets parallel and distributed applications and middleware. Libfabric is designed to minimize the impedance mismatch between applications, middleware and fabric communication hardware. Its interfaces target high- bandwidth, low-latency NICs, with a goal to scale to tens of thousands of nodes. Libfabric is supported by a variety of open source HPC middleware applications, including MPICH, Open MPI, Sandia SHMEM, Open SHMEM, Charm++, GasNET, Clang, UPC, DAOS, and others.
Libfabric source packages are available at: https://github.com/ofiwg/libfabric/releases
Please contact press@openfabrics.org with any questions.
