[R-package] allow use of MPI for distributed training

[esvhd]

Summary

As @jameslamb suggested in his very helpful replies to #3354 , adding a switch in build_r.R script to enable GPU and MPI would be very helpful.

Motivation

Enable MPI support for R packages. A switch is already available for python builds.

I'll let more capable minds to comment on the pros and cons for MPI vs sockets, and particularly when it comes to training.

Description

Current documentation for R package build might be missing some small details on how to enable MPI. See discussion in #3354 .
@jameslamb went through this and successfully managed to build the package with both GPU and MPI switched on.

[jameslamb]

Thanks for reporting this! I'll close this for now and add it to #2302 , where we keep all feature requests. Anyone reading this, you're welcome to add this feature!

I also want to add some additional information. The goal of this feature is to provide an experience like:

Rscript build_r.R --use-mpi

That should add -DUSE_MPI=ON to cmake_flags in R-package/src/install.libs.R.

Someone picking this up might also want to explore how this could be used for the CRAN package. A configure argument could be added so people can do something like this.

sh build-cran-package.sh
R CMD INSTALL lightgbm_3.0.0.tar.gz --configure-args='--use-mpi'

You can see more information on how to update the CRAN package at https://github.com/microsoft/LightGBM/blob/master/R-package/README.md#changing-the-cran-package. But getting this working for the CRAN package is not required.

[StrikerRUS]

I might be wrong, but I cannot find any references from R codebase to LGBM_NetworkInit() cpp function. If that's true, it means exposing --use-mpi flag will be not enough because actually R-package lacks support of parallel learning (even pure socket one). To enable it something similar to

LightGBM/python-package/lightgbm/basic.py

Lines 2347 to 2350 in 9eeac3c

	_safe_call(_LIB.LGBM_NetworkInit(c_str(machines),
	ctypes.c_int(local_listen_port),
	ctypes.c_int(listen_time_out),
	ctypes.c_int(num_machines)))

should be done.

[jameslamb]

you're right.