scikit-learn
diff --git a/‎doc/whats_new/v1.3.rst
Copy file name to clipboardExpand all lines: doc/whats_new/v1.3.rst
+11Lines changed: 11 additions & 0 deletions b/‎doc/whats_new/v1.3.rst
Copy file name to clipboardExpand all lines: doc/whats_new/v1.3.rst
+11Lines changed: 11 additions & 0 deletions
diff --git a/‎sklearn/utils/_openmp_helpers.pyx
Copy file name to clipboardExpand all lines: sklearn/utils/_openmp_helpers.pyx
+22-5Lines changed: 22 additions & 5 deletions b/‎sklearn/utils/_openmp_helpers.pyx
Copy file name to clipboardExpand all lines: sklearn/utils/_openmp_helpers.pyx
+22-5Lines changed: 22 additions & 5 deletions
@@ -126,6 +126,17 @@ Changes impacting all modules
 
   :pr:`25044` by :user:`Julien Jerphanion <jjerphan>`.
 
+- |Enhancement| All estimators that internally rely on OpenMP multi-threading
+  (via Cython) now use a number of threads equal to the number of physical
+  (instead of logical) cores by default. In the past, we observed that using as
+  many threads as logical cores on SMT hosts could sometimes cause severe
+  performance problems depending on the algorithms and the shape of the data.
+  Note that it is still possible to manually adjust the number of threads used
+  by OpenMP as documented in :ref:`parallelism`.
+
+  :pr:`26082` by :user:`Jérémie du Boisberranger <jeremiedbb>` and
+  :user:`Olivier Grisel <ogrisel>`.
+
 Changelog
 ---------
 
 
@@ -2,6 +2,12 @@ import os
 from joblib import cpu_count
 
 
+# Module level cache for cpu_count as we do not expect this to change during
+# the lifecycle of a Python program. This dictionary is keyed by
+# only_physical_cores.
+_CPU_COUNTS = {}
+
+
 def _openmp_parallelism_enabled():
     """Determines whether scikit-learn has been built with OpenMP
 
@@ -12,7 +18,7 @@ def _openmp_parallelism_enabled():
     return SKLEARN_OPENMP_PARALLELISM_ENABLED
 
 
-cpdef _openmp_effective_n_threads(n_threads=None, only_physical_cores=False):
+cpdef _openmp_effective_n_threads(n_threads=None, only_physical_cores=True):
     """Determine the effective number of threads to be used for OpenMP calls
 
     - For ``n_threads = None``,
@@ -33,6 +39,15 @@ cpdef _openmp_effective_n_threads(n_threads=None, only_physical_cores=False):
 
     - Raise a ValueError for ``n_threads = 0``.
 
+    Passing the `only_physical_cores=False` flag makes it possible to use extra
+    threads for SMT/HyperThreading logical cores. It has been empirically
+    observed that using as many threads as available SMT cores can slightly
+    improve the performance in some cases, but can severely degrade
+    performance other times. Therefore it is recommended to use
+    `only_physical_cores=True` unless an empirical study has been conducted to
+    assess the impact of SMT on a case-by-case basis (using various input data
+    shapes, in particular small data shapes).
+
     If scikit-learn is built without OpenMP support, always return 1.
     """
     if n_threads == 0:
@@ -47,10 +62,12 @@ cpdef _openmp_effective_n_threads(n_threads=None, only_physical_cores=False):
         # to exceed the number of cpus.
         max_n_threads = omp_get_max_threads()
     else:
-        max_n_threads = min(
-            omp_get_max_threads(),
-            cpu_count(only_physical_cores=only_physical_cores)
-        )
+        try:
+            n_cpus = _CPU_COUNTS[only_physical_cores]
+        except KeyError:
+            n_cpus = cpu_count(only_physical_cores=only_physical_cores)
+            _CPU_COUNTS[only_physical_cores] = n_cpus
+        max_n_threads = min(omp_get_max_threads(), n_cpus)
 
     if n_threads is None:
         return max_n_threads