AWS OFI NCCL is a plug-in which enables EC2 developers to use libfabric as a network provider while running NVIDIA's NCCL based applications.
Machine learning frameworks running on top of NVIDIA GPUs use a library called NCCL which provides standard collective communication routines for an arbitrary number of GPUs installed across single or multiple nodes.
This project implements a plug-in which maps NCCLs connection-oriented transport APIs to libfabric's connection-less reliable interface. This allows NCCL applications to take benefit of libfabric's transport layer services like reliable message support and operating system bypass.
The best way to build the plugin is to start with the latest release package. The plugin developers highly discourage customers from building directly from the HEAD of a GitHub branch, as releases go through more extensive testing than the pre-commit testing on git branches. More information about installing the plugin from a released tarball can be found in INSTALL.md.
Version numbers that end in -aws have only been tested on Amazon Web
Services Elastic Compute Cloud (EC2) instances and the Elastic Fabric
Adapter (EFA) network transport. Customers using other networks may
experience unexpected issues with these releases, but we welcome bug
reports if that is the case.
The plugin is regularly tested on the following operating systems:
Other operating systems are likely to work; there is very little distribution-specific code in the plugin.
To build the plugin, you need to have Libfabric and HWLOC installed prior to building the plugin., If you want to run the included multi-node tests, you also need an MPI Implementation installed. Each release of the plugin has a list of dependency versions in the top-level README.md file.
The plugin does not require NCCL to be pre-installed, but obviously a NCCL installation is required to use the plugin. As of NCCL 2.4.8, it is possible to use the same plugin build across multiple versions of NCCL (such as those installed per-package with Conda-like environments).
Most Libfabric providers should work with the plugin, possibly through
a utility provider. The plugin generally requires Reliable datagram
endpoints (FI_EP_RDM) with tagged messaging (FI_TAGGED, FI_MSG).
This is similar to the requirements of most MPI implementations and a
generally tested path in Libfabric. For GPUDirect RDMA support, the
plugin also requires FI_HMEM support, as well as RDMA support.
If you have any issues in building or using the package or if you think you may have found a bug, please open an issue.
Reporting issues and sending pull requests are always welcome. To learn how you can contribute, please look at our contributing guidelines.
This library is licensed under the Apache 2.0 License.