Deprecate single node multi-gpu mode (#4579)

* deprecate multi-gpu training

* add single node

* add warning

CMakeLists.txt

@@ -28,7 +28,7 @@ option(USE_NVTX "Build with cuda profiling annotations. Developers only." OFF)
 set(NVTX_HEADER_DIR "" CACHE PATH "Path to the stand-alone nvtx header")
 ## CUDA
 option(USE_CUDA  "Build with GPU acceleration" OFF)
-option(USE_NCCL  "Build with NCCL to enable multi-GPU support." OFF)
+option(USE_NCCL  "Build with NCCL to enable distributed GPU support." OFF)
 option(BUILD_WITH_SHARED_NCCL "Build with shared NCCL library." OFF)
 set(GPU_COMPUTE_VER "" CACHE STRING
   "Semicolon separated list of compute versions to be built against, e.g. '35;61'")

doc/build.rst

@@ -196,9 +196,9 @@ From the command line on Linux starting from the XGBoost directory:
   cmake .. -DUSE_CUDA=ON
   make -j4
 
-.. note:: Enabling multi-GPU training
+.. note:: Enabling distributed GPU training
 
-  By default, multi-GPU training is disabled and only a single GPU will be used. To enable multi-GPU training, set the option ``USE_NCCL=ON``. Multi-GPU training depends on NCCL2, available at `this link <https://developer.nvidia.com/nccl>`_. Since NCCL2 is only available for Linux machines, **multi-GPU training is available only for Linux**.
+  By default, distributed GPU training is disabled and only a single GPU will be used. To enable distributed GPU training, set the option ``USE_NCCL=ON``. Distributed GPU training depends on NCCL2, available at `this link <https://developer.nvidia.com/nccl>`_. Since NCCL2 is only available for Linux machines, **distributed GPU training is available only for Linux**.
 
   .. code-block:: bash
 

doc/gpu/index.rst

@@ -80,6 +80,8 @@ The GPU algorithms currently work with CLI, Python and R packages. See :doc:`/bu
 
 Single Node Multi-GPU
 =====================
+.. note:: Single node multi-GPU training is deprecated. Please use distributed GPU training with one process per GPU.
+
 Multiple GPUs can be used with the ``gpu_hist`` tree method using the ``n_gpus`` parameter. which defaults to 1. If this is set to -1 all available GPUs will be used.  If ``gpu_id`` is specified as non-zero, the selected gpu devices will be from ``gpu_id`` to ``gpu_id+n_gpus``, please note that ``gpu_id+n_gpus`` must be less than or equal to the number of available GPUs on your system.  As with GPU vs. CPU, multi-GPU will not always be faster than a single GPU due to PCI bus bandwidth that can limit performance.
 
 .. note:: Enabling multi-GPU training

include/xgboost/generic_parameters.h

@@ -66,7 +66,9 @@ struct LearnerTrainParam : public dmlc::Parameter<LearnerTrainParam> {
     DMLC_DECLARE_FIELD(n_gpus)
         .set_default(0)
         .set_lower_bound(-1)
-        .describe("Number of GPUs to use for multi-gpu algorithms.");
+        .describe("Deprecated, please use distributed training with one "
+                  "process per GPU. "
+                  "Number of GPUs to use for multi-gpu algorithms.");
     DMLC_DECLARE_FIELD(booster)
         .set_default("gbtree")
         .describe("Gradient booster used for training.");

src/learner.cc

@@ -200,8 +200,14 @@ class LearnerImpl : public Learner {
         << " Internal Error: Always call InitModel or Load before any evaluation.";
     this->ValidateDMatrix(dmat);
     CHECK(this->gbm_) << " Internal: GBM is not set";
-    if (this->gbm_->UseGPU() && cfg_.find("n_gpus") == cfg_.cend()) {
-      tparam_.n_gpus = 1;
+    if (this->gbm_->UseGPU()) {
+      if (cfg_.find("n_gpus") == cfg_.cend()) {
+        tparam_.n_gpus = 1;
+      }
+      if (tparam_.n_gpus != 1) {
+        LOG(WARNING) << "Multi-GPU training is deprecated. "
+                        "Please use distributed GPU training with one process per GPU.";
+      }
     }
   }