UCX Integration¶

Communication can be a major bottleneck in distributed systems. Dask-CUDA addresses this by supporting integration with UCX, an optimized communication framework that provides high-performance networking and supports a variety of transport methods, including NVLink and Infiniband for systems with specialized hardware, and TCP for systems without it. This integration is enabled through UCX-Py, an interface that provides Python bindings for UCX.

Requirements¶

Hardware¶

To use UCX with NVLink or InfiniBand, relevant GPUs must be connected with NVLink bridges or NVIDIA Mellanox InfiniBand Adapters, respectively. NVIDIA provides comparison charts for both NVLink bridges and InfiniBand adapters.

Software¶

UCX integration requires an environment with both UCX and UCX-Py installed; see UCX-Py Installation for detailed instructions on this process.

When using UCX, each NVLink and InfiniBand memory buffer must create a mapping between each unique pair of processes they are transferred across; this can be quite costly, potentially in the range of hundreds of milliseconds per mapping. For this reason, it is strongly recommended to use RAPIDS Memory Manager (RMM) to allocate a memory pool that is only prone to a single mapping operation, which all subsequent transfers may rely upon. A memory pool also prevents the Dask scheduler from deserializing CUDA data, which will cause a crash.

Configuration¶

In addition to installations of UCX and UCX-Py on your system, several options must be specified within your Dask configuration to enable the integration. Typically, these will affect UCX_TLS and UCX_SOCKADDR_TLS_PRIORITY, environment variables used by UCX to decide what transport methods to use and which to prioritize, respectively. However, some will affect related libraries, such as RMM:

ucx.cuda_copy: true – required.

Adds cuda_copy to UCX_TLS, enabling CUDA transfers over UCX.
ucx.tcp: true – required.

Adds tcp to UCX_TLS, enabling TCP transfers over UCX; this is required for very small transfers which are inefficient for NVLink and InfiniBand.
ucx.nvlink: true – required for NVLink.

Adds cuda_ipc to UCX_TLS, enabling NVLink transfers over UCX; affects intra-node communication only.
ucx.infiniband: true – required for InfiniBand.

Adds rc to UCX_TLS, enabling InfiniBand transfers over UCX.
ucx.rdmacm: true – recommended for InfiniBand.

Replaces sockcm with rdmacm in UCX_TLS and UCX_SOCKADDR_TLS_PRIORITY, enabling remote direct memory access (RDMA) for InfiniBand transfers. This is recommended by UCX for use with InfiniBand, and will not work if InfiniBand is disabled.
ucx.net-devices: <str> – recommended.

Explicitly sets UCX_NET_DEVICES instead of defaulting to "all", which can result in suboptimal performance. If using InfiniBand, set to "auto" to automatically detect the InfiniBand interface closest to each GPU. If InfiniBand is disabled, set to a UCX-compatible ethernet interface, e.g. "enp1s0f0" on a DGX-1. All available UCX-compatible interfaces can be listed by running ucx_info -d.

Warning

Setting ucx.net-devices: "auto" assumes that all InfiniBand interfaces on the system are connected and properly configured; undefined behavior may occur otherwise.
rmm.pool-size: <str|int> – recommended.

Allocates an RMM pool of the specified size for the process; size can be provided with an integer number of bytes or in human readable format, e.g. "4GB". It is recommended to set the pool size to at least the minimum amount of memory used by the process; if possible, one can map all GPU memory to a single pool, to be utilized for the lifetime of the process.

Note

These options can be used with mainline Dask/Distributed. However, some features are exclusive to Dask-CUDA, such as the automatic detection of InfiniBand interfaces. See Specializations for GPU Usage for more details on the benefits of using Dask-CUDA.

Usage¶

Dask-CUDA workers using UCX communication can be started manually with the dask-cuda-worker CLI tool or automatically with LocalCUDACluster. In either case, a dask.distributed.Client must be made for the worker cluster using the same UCX configuration.

dask-cuda-worker¶

A Dask cluster with UCX support can be started using the dask-cuda-worker CLI tool with a Dask scheduler which has been configured for UCX. This must be used for cases where a multi-node cluster is needed, as LocalCUDACluster will only start single-node clusters.

Scheduler¶

UCX configuration options will need to be specified for dask-scheduler as environment variables; see Dask Configuration for more details on the mapping between environment variables and options.

To start a Dask scheduler using UCX with all supported transports and a 1 gigabyte RMM pool:

DASK_UCX__CUDA_COPY=True \
DASK_UCX__TCP=True \
DASK_UCX__NVLINK=True \
DASK_UCX__INFINIBAND=True \
DASK_UCX__RDMACM=True \
DASK_UCX__NET_DEVICES=mlx5_0:1 \
DASK_RMM__POOL_SIZE=1GB \
dask-scheduler --protocol ucx --interface ib0

Note the specification of mlx5_0:1 as our UCX net device; because the scheduler does not rely upon Dask-CUDA, it cannot automatically detect InfiniBand interfaces, so we must specify one explicitly. We communicate to the scheduler that we will be using UCX with the --protocol option, and that we will be using InfiniBand with the --interface option.

To start the same Dask scheduler as above but only using NVLink:

DASK_UCX__CUDA_COPY=True \
DASK_UCX__TCP=True \
DASK_UCX__NVLINK=True \
DASK_RMM__POOL_SIZE=1GB \
dask-scheduler --protocol ucx --interface eth0

Note that we no longer specify a net device, as this generally can be skipped when using a non-InfiniBand interface.

Workers¶

All the relevant Dask configuration options for UCX have analogous parameters in dask-cuda-worker; see Worker for a complete list of these options.

To start workers with all supported transports and a 1 gigabyte RMM pool:

dask-cuda-worker ucx://<scheduler_address>:8786 \
--enable-tcp-over-ucx \
--enable-nvlink \
--enable-infiniband \
--enable-rdmacm \
--net-devices="auto" \
--rmm-pool-size="1GB"

Client¶

The UCX configurations used by the scheduler and client must be the same. This can be ensured by using dask_cuda.initialize, a utility which takes the same UCX configuring arguments as LocalCUDACluster and adds them to the current Dask configuration used when creating the client; see the API reference for a complete list of arguments.

To connect a client to a cluster with all supported transports:

from dask.distributed import Client
from dask_cuda.initialize import initialize

initialize(
    enable_tcp_over_ucx=True,
    enable_nvlink=True,
    enable_infiniband=True,
    enable_rdmacm=True,
    net_devices="mlx5_0:1",
)
client = Client("ucx://<scheduler_address>:8786")

Note the specification of "mlx5_0:1" as our net device; because the scheduler and client do not rely upon Dask-CUDA, they cannot automatically detect InfiniBand interfaces, so we must specify one explicitly.

LocalCUDACluster¶

All options available to dask-cuda-worker are also available as arguments for LocalCUDACluster; see the API reference for a complete list of arguments. When creating a LocalCUDACluster, dask_cuda.initialize is run automatically to ensure the Dask configuration is consistent with the cluster, so that a client can be connected to the cluster with no additional setup.

To start a cluster and client with all supported transports and a 1 gigabyte RMM pool:

from dask.distributed import Client
from dask_cuda import LocalCUDACluster

cluster = LocalCUDACluster(
    protocol="ucx",
    interface="ib0", # passed to the scheduler
    enable_tcp_over_ucx=True,
    enable_nvlink=True,
    enable_infiniband=True,
    enable_rdmacm=True,
    ucx_net_devices="auto",
    rmm_pool_size="1GB"
)
client = Client(cluster)