This is the communication library for BytePS. It is designed for high performance RDMA. However, it also supports TCP.

Build

git clone -b byteps https://github.com/bytedance/ps-lite
cd ps-lite 
make -j USE_RDMA=1

• Remove USE_RDMA=1 if you don't want to build with RDMA ibverbs support.
• Add USE_FABRIC=1 if you want to build with RDMA libfabric support for AWS Elastic Fabric Adaptor.

To build ps-lite with UCX:

# dependencies
sudo apt install -y build-essential libtool autoconf automake libnuma-dev unzip pkg-config

# build ucx
wget https://github.com/openucx/ucx/archive/refs/tags/v1.11.1.tar.gz
tar -xf v1.11.1.tar.gz
cd ucx-1.11.1
(./autogen.sh || ./autogen.sh) && ./configure --enable-logging --enable-mt --with-verbs --with-rdmacm --with-cuda=/usr/local/cuda
make clean && make -j && sudo make install -j

# build ps-lite
cd ..
make clean; USE_UCX=1 CUDA_HOME=/usr/local/cuda USE_CUDA=1 make -j

BytePS relies on UCXVan for GPU related communication, such as intra-node cuda-IPC, inter-node GPU-to-GPU / GPU-to-CPU communication with GPU-direct RDMA. For the list of transports UCX supports, see link.

Concepts

In ps-lite, there are three roles: worker, server and scheduler. Each role is an independent process.

The scheduler is responsible for setting up the connections between workers and servers at initialization. There should be only 1 scheduler process.

A worker process only communicates with server processes, and vice versa. There won't be any traffic between worker-to-worker, and server-to-server.

Tutorial

After build, you will have two testing applications under tests/ dir, namely test_benchmark and test_ipc_benchmark. Below we elaborate how you can run with them.

To debug, set PS_VERBOSE=1 to see important logs during connection setup, and PS_VERBOSE=2 to see each message log.

1. Basic benchmark

Suppose you want to run with 1 worker and 1 server on different machines. Therefore, we need to launch 3 processes in total (including the scheduler). You can launch the scheduler process at any machine as it does not affect the performance.

For the scheduler:

# common setup
export DMLC_ENABLE_RDMA=ibverbs
export DMLC_NUM_WORKER=1
export DMLC_NUM_SERVER=1 
export DMLC_PS_ROOT_URI=10.0.0.2  # scheduler's RDMA interface IP 
export DMLC_PS_ROOT_PORT=8123     # scheduler's port (can random choose)
export DMLC_INTERFACE=eth5        # my RDMA interface 

# launch scheduler
DMLC_ROLE=scheduler ./tests/test_benchmark

For the server:

# common setup
export DMLC_ENABLE_RDMA=ibverbs
export DMLC_NUM_WORKER=1
export DMLC_NUM_SERVER=1 
export DMLC_PS_ROOT_URI=10.0.0.2  # scheduler's RDMA interface IP 
export DMLC_PS_ROOT_PORT=8123     # scheduler's port (can random choose)
export DMLC_INTERFACE=eth5        # my RDMA interface 

# launch server
DMLC_ROLE=server ./tests/test_benchmark

For the worker:

# common setup
export DMLC_ENABLE_RDMA=ibverbs
export DMLC_NUM_WORKER=1
export DMLC_NUM_SERVER=1 
export DMLC_PS_ROOT_URI=10.0.0.2  # scheduler's RDMA interface IP 
export DMLC_PS_ROOT_PORT=8123     # scheduler's port (can random choose)
export DMLC_INTERFACE=eth5        # my RDMA interface 

# launch worker
DMLC_ROLE=worker ./tests/test_benchmark

If you want to use libfabric with Amazon Elastic Fabric Adaptor, make sure to set DMLC_ENABLE_RDMA=fabric for all processes. If you are using libfabric < 1.10, please also set FI_EFA_ENABLE_SHM_TRANSFER=0 to avoid a bug in the EFA shm provider.

If you just want to use TCP, make sure to unset DMLC_ENABLE_RDMA for all processes.

2. Benchmark with IPC support

The test_ipc_benchmark demonstrates how inter-process communication (IPC) helps improve RDMA performance when the server is co-located with the worker.

Suppose you have two machines. Each machine should launch a worker and a server process.

For the scheduler: (you can launch it on either machine-0 or machine-1)

# common setup
export DMLC_ENABLE_RDMA=ibverbs
export DMLC_NUM_WORKER=2
export DMLC_NUM_SERVER=2 
export DMLC_PS_ROOT_URI=10.0.0.2  # scheduler's RDMA interface IP 
export DMLC_PS_ROOT_PORT=8123     # scheduler's port (can random choose)
export DMLC_INTERFACE=eth5        # my RDMA interface 

# launch scheduler
DMLC_ROLE=scheduler ./tests/test_ipc_benchmark

For machine-0 and machine-1:

# common setup
export DMLC_ENABLE_RDMA=ibverbs
export DMLC_NUM_WORKER=2
export DMLC_NUM_SERVER=2 
export DMLC_PS_ROOT_URI=10.0.0.2  # scheduler's RDMA interface IP 
export DMLC_PS_ROOT_PORT=8123     # scheduler's port (can random choose)
export DMLC_INTERFACE=eth5        # my RDMA interface 

# launch server and worker
DMLC_ROLE=server ./tests/test_ipc_benchmark &
DMLC_ROLE=worker ./tests/test_ipc_benchmark 

Note: This benchmark is only valid for RDMA.

3. Other GPU-related benchmarks

cd tests;
NODE_ONE_IP=xxx NODE_TWO_IP=yyy bash test.sh (local|remote|joint) bytes_per_msg msg_count (push_only|pull_only|push_pull) (cpu2cpu|cpu2gpu|gpu2gpu|gpu2cpu)