Throw again after logging that RMM could not intialize

[abellina]

Signed-off-by: Alessandro Bellina abellina@nvidia.com

Closes: #5242.

This is a (small) proposed diff to make failure to initialize RMM (which throws a CudfException) fatal, instead of silently continuing with a default initialized raw cuda memory resource.

The executor logs will look like this:

22/04/13 14:56:06 ERROR GpuDeviceManager: Could not initialize RMM, exiting!
 ai.rapids.cudf.CudfException: RMM failure at: /home/jenkins/agent/workspace/jenkins-cudf-for-dev-32-cuda11/cpp/build/_deps/rmm-src/include/rmm/mr/device/    cuda_async_memory_resource.hpp:67: cudaMallocAsync not supported with this CUDA driver/runtime version
  at ai.rapids.cudf.Rmm.initializeInternal(Native Method)

[revans2]

This is typically an anti-pattern

https://stackoverflow.com/questions/6639963/why-is-log-and-throw-considered-an-anti-pattern

Is there a reason we need the log statement? Should we just lat the exception continue up?

[abellina]

yes agree, I wanted to add something more specific than the error seen here: #5242 (comment), that's the only reason.

Currently this exception that I am re-throwing gets caught up again in the RapidsExecutorPlugin where we exit. Perhaps I should change the message here to say we are exiting. Do you have a preference?

22/04/13 14:39:05 ERROR RapidsExecutorPlugin: Exception in the executor plugin

[revans2]

I would rather have the code that exits say we are exiting, otherwise there is coupling between the code that exits and here for no good reason.

[abellina]

@revans2 handled here: f6545d3

[abellina]

build