Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Smarter choice of histogram construction for distributed gpu_hist (#4519)

Browse Source 
* Smarter choice of histogram construction for distributed gpu_hist

* Limit omp team size in ExecuteShards
This commit is contained in:
Rory Mitchell
2019-05-31 14:11:34 +12:00
committed by GitHub
parent dd60fc23e6
commit fbbae3386a
3 changed files with 67 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/tree/updater_gpu_hist.cu
View File

@@ -1054,15 +1054,17 @@ struct DeviceShard {
auto build_hist_nidx = nidx_left;
auto subtraction_trick_nidx = nidx_right;
// If we are using a single GPU, build the histogram for the node with the
// fewest training instances
// If we are distributed, don't bother
if (reducer->IsSingleGPU()) {
bool fewer_right =
ridx_segments[nidx_right].Size() < ridx_segments[nidx_left].Size();
if (fewer_right) {
std::swap(build_hist_nidx, subtraction_trick_nidx);
}
auto left_node_rows = ridx_segments[nidx_left].Size();
auto right_node_rows = ridx_segments[nidx_right].Size();
// Decide whether to build the left histogram or right histogram
// Find the largest number of training instances on any given Shard
// Assume this will be the bottleneck and avoid building this node if
// possible
std::vector<size_t> max_reduce = {left_node_rows, right_node_rows};
reducer->HostMaxAllReduce(&max_reduce);
bool fewer_right = max_reduce[1] < max_reduce[0];
if (fewer_right) {
std::swap(build_hist_nidx, subtraction_trick_nidx);
}
this->BuildHist(build_hist_nidx);