set env vars for xgboost

[sebffischer]

Sys.setenv(OMP_NUM_THREADS=1) Sys.setenv(OMP_THREAD_LIMIT=1)

[be-marc]

The nthread parameter works as expected.

library(mlr3learners)

task = tsk("spam")

learner = lrn("classif.xgboost", nrounds = 2000)
system.time(learner$train(task))
#    user  system elapsed
#  24.984   0.024  20.503

learner = lrn("classif.xgboost", nrounds = 2000, nthread = 2)
system.time(learner$train(task))
#    user  system elapsed 
#  28.974   0.329  12.362 

Run on the lrz. Looks like there is no problem. The increase in training time can be explained by the fact that the training data set becomes larger when there are more folds. The longer total runtime is due to future.

library(mlr3learners)

task = tsk("spam")

learner = lrn("classif.xgboost", nrounds = 2000)
system.time(learner$train(task))
#    user  system elapsed
#  24.984   0.024  20.503

future::plan("multisession", workers = 3)
system.time({rr = resample(task, learner, rsmp("cv", folds = 3))})
#  user  system elapsed
# 2.079   0.093  17.363

mean(mlr3misc::map_dbl(rr$learners, function(x) x$state$train_time))
# [1] 13.21067

future::plan("multisession", workers = 6)
system.time({rr = resample(task, learner, rsmp("cv", folds = 6))})
  #  user  system elapsed
  # 2.822   0.054  22.336

mean(mlr3misc::map_dbl(rr$learners, function(x) x$state$train_time))
# [1] 16.67617

future::plan("multisession", workers = 12)
system.time({rr = resample(task, learner, rsmp("cv", folds = 12))})
  #  user  system elapsed
  # 4.345   0.119  28.737

mean(mlr3misc::map_dbl(rr$learners, function(x) x$state$train_time))
# [1] 18.75075

future::plan("multisession", workers = 40)
system.time({rr = resample(task, learner, rsmp("cv", folds = 40))})
 #   user  system elapsed
 # 16.879   0.337  53.033

mean(mlr3misc::map_dbl(rr$learners, function(x) x$state$train_time))
# [1] 21.87785

The environment variables don't change the results.

[sebffischer]

Ok weird, I think Lukas reported that adding this to his .Rprofile improved stuff for him.

[be-marc]

Yes a year ago this also fixed my benchmark. Either this has since been fixed in xgboost or it was a weird data.table with xgboost combination after all. But good to know that our nthread parameter works.

[be-marc]

On my local machine it's different but this might due to high and low performance cores. The environment variables also don't change the results on my local machine.