Support for Multi-GPU XGBoost

post by william_schram on Mar 4, 2020

william_schram

Lambda,

I’m trying to max out the value of my workstation by transitioning from single-GPU XGBoost to multi-GPU. Can I convince you to support and/or write an install guide?

My attempts to make this work have led me down a path of installing NCCL2 from nvidia, upgrading cmake to version 3.16.4 (>3.12), and building XGBoost for GPU.

Where I am failing is I haven’t worked out the location of nccl (which must be specified in the compiling process of the build). Commands from XGBoost documentation summarized below:

mkdir build

cd build

cmake … -DUSE_CUDA=ON -DUSE_NCCL=ON -DNCCL_ROOT=/path/to/nccl2

make -j4

And finally, how do you install this build, or is it installed once it is built?

The documentation from XGBoost isn’t sufficient for this data scientist to follow. I’m hoping someone more familiar with building software with ‘make’ and ‘cmake’ can help.

My Lambda workstation is running Ubuntu 18.04 LTS, CUDA 10.2, nvidia driver 440.44.