NCCL Fast Socket is a transport layer plugin to improve NCCL collective communication performance on Google Cloud.

Collective communication primitives such as all-reduce and all-gather have been widely used in distributed training in machine learning. The NVIDIA Collective Communications Library (NCCL) is a highly optimized implementation of these multi-GPU and multi-node collective communication primitives that supports NVIDIA GPUs.

NCCL Fast Socket is based on TCP/IP communication and uses a number of techniques to achieve better and more consistent performance, especially with 100 Gbps networking on Google Cloud.

Fast Socket requires working installation of CUDA to build. After building the plugin, it has to be in LD_LIBRARY_PATH in order to be loaded by NCCL.

The plugin uses Bazel to build. You can build the plugin as follows:

$ bazel build :all

The plugin is located at bazel-bin/libnccl-net.so and can be copied into your LD_LIBRARY_PATH.

Please open an issue if you have any questions or if you think you may have found any bugs.

Contributions are always welcomed. Please refer to our contributing guidelines to learn how to contriute.

Fast Socket is licensed under the terms of a BSD-style license. See LICENSE for more information.

This is not an officially supported Google product.