Enabling UCX communication

A CUDA cluster using UCX communication can be started automatically with LocalCUDACluster or manually with the dask-cuda-worker CLI tool. In either case, a dask.distributed.Client must be made for the worker cluster using the same Dask UCX configuration; see UCX Integration – Configuration for details on all available options.

LocalCUDACluster with Automatic Configuration

Automatic configuration was introduced in Dask-CUDA 22.02 and requires UCX >= 1.11.1. This allows the user to specify only the UCX protocol and let UCX decide which transports to use.

To connect a client to a cluster with automatically-configured UCX and an RMM pool:

from dask.distributed import Client
from dask_cuda import LocalCUDACluster

cluster = LocalCUDACluster(
    protocol="ucx",
    interface="ib0",
    rmm_pool_size="1GB"
)
client = Client(cluster)

Note

The interface="ib0" is intentionally specified above to ensure RDMACM is used in systems that support InfiniBand. On systems that don’t support InfiniBand or where RDMACM isn’t required, the interface argument may be omitted or specified to listen on a different interface.

LocalCUDACluster with Manual Configuration

When using LocalCUDACluster with UCX communication and manual configuration, all required UCX configuration is handled through arguments supplied at construction; see API – Cluster for a complete list of these arguments. To connect a client to a cluster with all supported transports and an RMM pool:

from dask.distributed import Client
from dask_cuda import LocalCUDACluster

cluster = LocalCUDACluster(
    protocol="ucx",
    interface="ib0",
    enable_tcp_over_ucx=True,
    enable_nvlink=True,
    enable_infiniband=True,
    enable_rdmacm=True,
    rmm_pool_size="1GB"
)
client = Client(cluster)

dask-cuda-worker with Automatic Configuration

When using dask-cuda-worker with UCX communication and automatic configuration, the scheduler, workers, and client must all be started manually, but without specifying any UCX transports explicitly. This is only supported in Dask-CUDA 22.02 and newer and requires UCX >= 1.11.1.

Scheduler

For automatic UCX configuration, we must ensure a CUDA context is created on the scheduler before UCX is initialized. This can be satisfied by specifying the DASK_DISTRIBUTED__COMM__UCX__CREATE_CUDA_CONTEXT=True environment variable when creating the scheduler.

To start a Dask scheduler using UCX with automatic configuration and one GB of RMM pool:

$ DASK_DISTRIBUTED__COMM__UCX__CREATE_CUDA_CONTEXT=True \
> DASK_DISTRIBUTED__RMM__POOL_SIZE=1GB \
> dask-scheduler --protocol ucx --interface ib0

Note

The interface="ib0" is intentionally specified above to ensure RDMACM is used in systems that support InfiniBand. On systems that don’t support InfiniBand or where RDMACM isn’t required, the interface argument may be omitted or specified to listen on a different interface.

We specify UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda above for optimal performance with InfiniBand, see details here. If not using InfiniBand, that option may be omitted. In UCX 1.12 and newer, that option is default and may be omitted as well even when using InfiniBand.

Workers

To start workers with automatic UCX configuration and an RMM pool of 14GB per GPU:

$ UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda
> dask-cuda-worker ucx://<scheduler_address>:8786 \
> --rmm-pool-size="14GB" \
> --interface="ib0"

Note

Analogous to the scheduler setup, the interface="ib0" is intentionally specified above to ensure RDMACM is used in systems that support InfiniBand. On systems that don’t support InfiniBand or where RDMACM isn’t required, the interface argument may be omitted or specified to listen on a different interface.

We specify UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda above for optimal performance with InfiniBand, see details here. If not using InfiniBand, that option may be omitted. In UCX 1.12 and newer, that option is default and may be omitted as well even when using InfiniBand.

Client

To connect a client to the cluster with automatic UCX configuration we started:

import os

os.environ["UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES"] = "cuda"

import dask
from dask.distributed import Client

with dask.config.set({"distributed.comm.ucx.create_cuda_context": True}):
    client = Client("ucx://<scheduler_address>:8786")

Alternatively, the with dask.config.set statement from the example above may be omitted and the DASK_DISTRIBUTED__COMM__UCX__CREATE_CUDA_CONTEXT=True environment variable specified instead:

import os

os.environ["UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES"] = "cuda"
os.environ["DASK_DISTRIBUTED__COMM__UCX__CREATE_CUDA_CONTEXT"] = "True"

from dask.distributed import Client

client = Client("ucx://<scheduler_address>:8786")

Note

We specify UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda above for optimal performance with InfiniBand, see details here. If not using InfiniBand, that option may be omitted. In UCX 1.12 and newer, that option is default and may be omitted as well even when using InfiniBand.

dask-cuda-worker with Manual Configuration

When using dask-cuda-worker with UCX communication and manual configuration, the scheduler, workers, and client must all be started manually, each using the same UCX configuration.

Scheduler

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

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

$ DASK_DISTRIBUTED__COMM__UCX__CUDA_COPY=True \
> DASK_DISTRIBUTED__COMM__UCX__TCP=True \
> DASK_DISTRIBUTED__COMM__UCX__NVLINK=True \
> DASK_DISTRIBUTED__COMM__UCX__INFINIBAND=True \
> DASK_DISTRIBUTED__COMM__UCX__RDMACM=True \
> DASK_DISTRIBUTED__RMM__POOL_SIZE=1GB \
> dask-scheduler --protocol ucx --interface ib0

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.

Workers

All UCX configuration options have analogous options in dask-cuda-worker; see API – Worker for a complete list of these options. To start a cluster with all supported transports and an RMM pool:

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

Client

A client can be configured to use UCX 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 it; see API – Client initialization for a complete list of arguments. To connect a client to the cluster we have made:

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,
)
client = Client("ucx://<scheduler_address>:8786")