[SPARK-28843] Set OMP_NUM_THREADS to executor cores reduce Python memory consumption

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.3.3, 2.4.3, 3.0.0
• Fix Version/s: 3.0.0
• Component/s: PySpark
• Labels: release-notes

Docs Text

Pyspark workers now set the env variable OMP_NUM_THREADS (if not already set) to the number of cores used by an executor (spark.executor.cores). When unset, it defaulted to the total number of VM cores. This avoids excessively large OpenMP thread pools when using, for example, numpy.

• Assignee: Ryan Blue
• Reporter: Ryan Blue
• Created: 21/Aug/19 22:54
• Updated: 28/Feb/23 06:21
• Resolved: 30/Aug/19 01:30

Description

While testing hardware with more cores, we found that the amount of memory required by PySpark applications increased and tracked the problem to importing numpy. The numpy issue is https://github.com/numpy/numpy/issues/10455

NumPy uses OpenMP that starts a thread pool with the number of cores on the machine (and does not respect cgroups). When we set this lower we see a significant reduction in memory consumption.

This parallelism setting should be set to the number of cores allocated to the executor, not the number of cores available.

Issue Links

• is duplicated by: SPARK-28846 Set OMP_NUM_THREADS to executor cores for python
• relates to: SPARK-42613 PythonRunner should set OMP_NUM_THREADS to task cpus instead of executor cores by default
• links to: GitHub Pull Request #25545