Handle OMP_THREAD_LIMIT. (#7390)

R-package/R/xgboost.R

@@ -9,8 +9,8 @@ xgboost <- function(data = NULL, label = NULL, missing = NA, weight = NULL,
                     early_stopping_rounds = NULL, maximize = NULL,
                     save_period = NULL, save_name = "xgboost.model",
                     xgb_model = NULL, callbacks = list(), ...) {
-
-  dtrain <- xgb.get.DMatrix(data, label, missing, weight, nthread = params$nthread)
+  merged <- check.booster.params(params, ...)
+  dtrain <- xgb.get.DMatrix(data, label, missing, weight, nthread = merged$nthread)
 
   watchlist <- list(train = dtrain)
 

src/common/threading_utils.h

@@ -7,11 +7,28 @@
 #define XGBOOST_COMMON_THREADING_UTILS_H_
 
 #include <dmlc/common.h>
-#include <vector>
+#include <dmlc/omp.h>
+
 #include <algorithm>
+#include <limits>
 #include <type_traits>  // std::is_signed
+#include <vector>
+
 #include "xgboost/logging.h"
 
+#if !defined(_OPENMP)
+extern "C" {
+inline int32_t omp_get_thread_limit() __GOMP_NOTHROW { return 1; }  // NOLINT
+}
+#endif  // !defined(_OPENMP)
+
+// MSVC doesn't implement the thread limit.
+#if defined(_OPENMP) && defined(_MSC_VER)
+extern "C" {
+inline int32_t omp_get_thread_limit() { return std::numeric_limits<int32_t>::max(); }  // NOLINT
+}
+#endif  // defined(_MSC_VER)
+
 namespace xgboost {
 namespace common {
 
@@ -153,7 +170,7 @@ struct Sched {
 };
 
 template <typename Index, typename Func>
-void ParallelFor(Index size, size_t n_threads, Sched sched, Func fn) {
+void ParallelFor(Index size, int32_t n_threads, Sched sched, Func fn) {
 #if defined(_MSC_VER)
   // msvc doesn't support unsigned integer as openmp index.
   using OmpInd = std::conditional_t<std::is_signed<Index>::value, Index, omp_ulong>;
@@ -220,6 +237,13 @@ void ParallelFor(Index size, size_t n_threads, Func fn) {
 template <typename Index, typename Func>
 void ParallelFor(Index size, Func fn) {
   ParallelFor(size, omp_get_max_threads(), Sched::Static(), fn);
+}                                        // !defined(_OPENMP)
+
+
+inline int32_t OmpGetThreadLimit() {
+  int32_t limit = omp_get_thread_limit();
+  CHECK_GE(limit, 1) << "Invalid thread limit for OpenMP.";
+  return limit;
 }
 
 /* \brief Configure parallel threads.
@@ -235,15 +259,18 @@ inline int32_t OmpSetNumThreads(int32_t* p_threads) {
   if (threads <= 0) {
     threads = omp_get_num_procs();
   }
+  threads = std::min(threads, OmpGetThreadLimit());
   omp_set_num_threads(threads);
   return nthread_original;
 }
+
 inline int32_t OmpSetNumThreadsWithoutHT(int32_t* p_threads) {
   auto& threads = *p_threads;
   int32_t nthread_original = omp_get_max_threads();
   if (threads <= 0) {
     threads = nthread_original;
   }
+  threads = std::min(threads, OmpGetThreadLimit());
   omp_set_num_threads(threads);
   return nthread_original;
 }
@@ -252,6 +279,7 @@ inline int32_t OmpGetNumThreads(int32_t n_threads) {
   if (n_threads <= 0) {
     n_threads = omp_get_num_procs();
   }
+  n_threads = std::min(n_threads, OmpGetThreadLimit());
   return n_threads;
 }
 }  // namespace common

tests/python/test_openmp.py

@@ -1,6 +1,12 @@
-# -*- coding: utf-8 -*-
+import os
+import tempfile
+import subprocess
+
 import xgboost as xgb
 import numpy as np
+import pytest
+
+import testing as tm
 
 
 class TestOMP:
@@ -71,3 +77,31 @@ class TestOMP:
         assert auc_1 == auc_2 == auc_3
         assert np.array_equal(auc_1, auc_2)
         assert np.array_equal(auc_1, auc_3)
+
+    @pytest.mark.skipif(**tm.no_sklearn())
+    def test_with_omp_thread_limit(self):
+        args = [
+            "python", os.path.join(
+                tm.PROJECT_ROOT, "tests", "python", "with_omp_limit.py"
+            )
+        ]
+        results = []
+        with tempfile.TemporaryDirectory() as tmpdir:
+            for i in (1, 2, 16):
+                path = os.path.join(tmpdir, str(i))
+                with open(path, "w") as fd:
+                    fd.write("\n")
+                cp = args.copy()
+                cp.append(path)
+
+                env = os.environ.copy()
+                env["OMP_THREAD_LIMIT"] = str(i)
+
+                status = subprocess.call(cp, env=env)
+                assert status == 0
+
+                with open(path, "r") as fd:
+                    results.append(float(fd.read()))
+
+        for auc in results:
+            np.testing.assert_allclose(auc, results[0])

tests/python/with_omp_limit.py

@@ -0,0 +1,26 @@
+import os
+import xgboost as xgb
+from sklearn.datasets import make_classification
+from sklearn.metrics import roc_auc_score
+import sys
+
+
+def run_omp(output_path: str):
+    X, y = make_classification(
+        n_samples=200, n_features=32, n_classes=3, n_informative=8
+    )
+    Xy = xgb.DMatrix(X, y, nthread=16)
+    booster = xgb.train(
+        {"num_class": 3, "objective": "multi:softprob", "n_jobs": 16},
+        Xy,
+        num_boost_round=8,
+    )
+    score = booster.predict(Xy)
+    auc = roc_auc_score(y, score, average="weighted", multi_class="ovr")
+    with open(output_path, "w") as fd:
+        fd.write(str(auc))
+
+
+if __name__ == "__main__":
+    out = sys.argv[1]
+    run_omp(out)