BytePS is a high performance and general distributed training framework. It supports TensorFlow, Keras, PyTorch, and MXNet, and can run on either TCP or RDMA network.
BytePS outperforms existing open-sourced distributed training frameworks by a large margin. For example, on a popular public cloud and with the same number of GPUs, BytePS can double the training speed (see below), compared with Horovod+NCCL.
For demonstration, we test two models: VGG16 (communication-intensive) and Resnet50 (computation-intensive). Both models are trained using fp32.
We use Tesla V100 16GB GPUs and set batch size equal to 64 per GPU. The machines are in fact VMs on a popular public cloud. Each machine has 8 V100 GPUs with NVLink-enabled. Machines are inter-connected with 20 Gbps TCP/IP network.
BytePS outperforms Horovod (NCCL) by 44% for Resnet50, and 100% for VGG16.
You can reproduce the results using the Dockerfiles and example scripts we provide.
Evaluation on RDMA networks can be found at performance.md.
How can BytePS outperform Horovod by so much? One of the main reasons is that BytePS is designed for cloud and shared clusters, and throws away MPI.
MPI was born in the HPC world and is good for a cluster built with homogeneous hardware and for running a single job. However, cloud (or in-house shared clusters) is different.
This leads us to rethink the best communication strategy, as explained in here. In short, BytePS only uses NCCL inside a machine, while re-implements the inter-machine communication.
BytePS also incorporates many acceleration techniques such as hierarchical strategy, pipelining, tensor partitioning, NUMA-aware local communication, priority-based scheduling, etc.
We provide a step-by-step tutorial for you to run benchmark training tasks. After you can start BytePS, read best practice to get the best performance.
Below, we explain how to build and run BytePS by yourself. BytePS assumes that you have already installed one or more of the following frameworks: TensorFlow / PyTorch / MXNet. BytePS depends on CUDA and NCCL, and requires gcc>=4.9. If you are working on CentOS/Redhat and have gcc<4.9, you can try yum install devtoolset-7 before everything else.
You can download our wheels and install. Please refer to pip-list.md for more instructions.
If the above does not contain your desired wheel resource, or you want to try building from source code:
git clone --recurse-submodules https://github.com/bytedance/byteps
cd byteps
python setup.py install
Notes:
- Please pin your gcc to 4.9 before building, here is an example.
- You may set
BYTEPS_USE_RDMA=1to install with RDMA support. Before this, make sure your RDMA drivers have been properly installed and tested.
For your server and scheduler node, we highly recommend you to just use our prebuilt docker image bytepsimage/byteps_server (TCP) or bytepsimage/byteps_server_rdma (RDMA). Otherwise, you have to manually compile our modified MXNet as in our dockerfiles: Dockerfile.server and Dockerfile.server.rdma.
Refer to Documentations for how to launch distributed jobs and more detailed configurations.
Though being totally different at its core, BytePS is highly compatible with Horovod interfaces (Thank you, Horovod community!). We chose Horovod interfaces in order to minimize your efforts for testing BytePS.
If your tasks only rely on Horovod's allreduce and broadcast, you should be able to switch to BytePS in 1 minute. Simply replace import horovod.tensorflow as hvd by import byteps.tensorflow as bps, and then replace all hvd in your code by bps. If your code invokes hvd.allreduce directly, you should also replace it by bps.push_pull.
Many of our examples were copied from Horovod and modified in this way. For instance, compare the MNIST example for BytePS and Horovod.
BytePS does not support pure CPU training for now. One reason is that the cheap PS assumption of BytePS do not hold for CPU training. Consequently, you need CUDA and NCCL to build and run BytePS.
We would like to have below features, and it is not hard to implement them in BytePS architecture. However, they are not implemented yet:
- Sparse model training
- Asynchronous training
- Fault-tolerance
- Straggler-mitigation