analysis of performance characteristics in Python's experimental free-threading support (PEP 703) for Python 3.13+.

• Thread Concepts (thread affinity, ref counting, false sharing and specialization)
• Memory Barriers (memory ordering, false sharing and padded memory)
• PEP 703: Making the GIL Optional
• Free Threading HOWTO

Establish baseline performance metrics:

• cpu-intensive performance
• mem-io performance
• thread creation

• dispatch overhead

• thread contention
• lock patterns

Memory Ordering:

• sequential access
• strided access
• memory alignment
• false sharing (baseline)
• false sharing (padded)

Reference Counting:

• thread-local objects
• cross-thread objects

• dynamic operations
• typed operations

Runners:

• Github actions runner:

  • CPU: 4 cores @ 2923.43 MHz
  • CPU affinity: All 4 cores
  • Memory: 15.61 GB
  • OS: Linux 6.5.0-1025-azure

• Local machine:

  • CPU: 20 cores @ 2022.67 MHz (Intel i7-14700K)
  • CPU affinity: 2-7 (6 isolated P-cores, refer to this article for details on isolation and cpu-pinning)
  • Memory: 62.57 GB
  • OS: Linux 6.8.0-47-generic

# Run all tests
python benchmark_runner.py

# Run specific category
python benchmark_runner.py --benchmarks memory/ordering/*

# Run with profiling
python benchmark_runner.py --profile detailed

• Linux kernel 5.10+
• Python versions:
  • 3.12.7 (baseline)
  • 3.13.0 (with GIL)
  • 3.13.0t (no GIL)

• benchmark-results

study more after the following areas are addressed in Python 3.14+:

1. Re-enabling specializing adaptive interpreter
2. Reduced immortalization scope
3. Improved thread-safety mechanisms
4. Better memory usage patterns