Skip to content

DEBUG log messages imply UCX config options are invalid #4845

@rlratzel

Description

@rlratzel

When debugging is enabled, distributed outputs the following messages:

distributed.comm.ucx - DEBUG - Key: cuda_copy with value: False not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: tcp with value: False not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: nvlink with value: False not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: infiniband with value: False not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: rdmacm with value: False not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: net-devices with value: None not a valid UCX configuration option.
distributed.comm.ucx - DEBUG - Key: reuse-endpoints with value: None not a valid UCX configuration option.

When the corresponding env vars as set, such as export DASK_UCX__CUDA_COPY=True, the code in ucx.py here seems to behave correctly, but the check here still fails and prints the debug messages.

This can be seen simply by setting:

export DASK_LOGGING__DISTRIBUTED="DEBUG"

and running the scheduler like so:

python -m distributed.cli.dask_scheduler

Since the default values even cause the debug messages to be printed, there's no need to set the UCX config options to see the debug messages above.

@quasiben @pentschev

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions