Communicate all lightgbm distributed metrics into node 0 for surfacing as aggregate

[Jeff Omhover (jfomhover)]

Problem

When using distributed lightgbm, each node will report its own validation metrics.

Proposed Design

Discarded alternative

An initial design here https://github.com/microsoft/lightgbm-benchmark/blob/c45eacfc284aff286c46ee9349455466ca07810f/src/common/lightgbm_utils.py consisted in using mpi naively during each of the callback calls (synchronously to each callback). This creates some interaction with LightGBM's internal MPI initialization somehow, and ended up causing the exception below on node 0:

  File "train.py", line 265, in run
    booster = lightgbm.train(
  File "/azureml-envs/lightgbm/lib/python3.8/site-packages/lightgbm/engine.py", line 293, in train
    booster.update(fobj=fobj)
  File "/azureml-envs/lightgbm/lib/python3.8/site-packages/lightgbm/basic.py", line 3021, in update
    _safe_call(_LIB.LGBM_BoosterUpdateOneIter(
  File "/azureml-envs/lightgbm/lib/python3.8/site-packages/lightgbm/basic.py", line 125, in _safe_call
    raise LightGBMError(_LIB.LGBM_GetLastError().decode('utf-8'))
lightgbm.basic.LightGBMError: Please initialize the network interface first

That's why we ended up using threading instead.

[github-actions]

Unit Test Results for Build

  1 files  ±0    1 suites  ±0   8s ⏱️ -1s
76 tests ±0  76 ✔️ ±0  0 💤 ±0  0 ❌ ±0 

Results for commit ce90af5. ± Comparison against base commit 7b8a339.