[FEA] Rapids Plugin is able to recognize the GPU device used in Executor and leave a record in Spark Eventlog

[wjxiz1992]

Is your feature request related to a problem? Please describe.
I wish the RAPIDS Accelerator for Apache Spark would be able to recognize the GPU device sku/type and is also able to write it in the spark eventlog.

Rational:

1. There may be some new features in the future that start to require such information for auto-tuning moves.
2. Leave a record in the eventlog so the spark-rapids-tools can get it and give out config suggestions based on the GPU type.

Describe the solution you'd like
No good idea yet as common cases are Driver nodes don't have GPU installed, and nvidia-smi must be called at executor. Maybe SparkEvent can help in this case.
Describe alternatives you've considered
/

Additional context
/

[mattahrens]

@revans2 -- would this be feasible along with planned zero config changes for concurrentGpuTasks?

[gerashegalov]

I think we should do this in JNI on the executors https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaDeviceProp.html#structcudaDeviceProp_11e26f1c6bd42f4821b7ef1a4bd3bd25c instead of nvidia -smi and expose it as a metric, Thus the driver can store an event with deviceName:count pairs