Skip to content

adr-002-thread-affinity.md

Latest commit

 

History

History
30 lines (24 loc) · 1023 Bytes

adr-002-thread-affinity.md

File metadata and controls

30 lines (24 loc) · 1023 Bytes
Raw

ADR-002: Thread Affinity — Not Needed

Context

P3-4 called for thread affinity investigation to reduce contention in multi-threaded PyTorch operations during CFD solves.

Decision

CLOSE: thread affinity is not needed. No further investment.

Evidence

Based on scripts/profile_results.json (2026-05-29):

  • Baseline (default threading): 0.1283s ± 0.0033s
  • Pinned thread: 0.1328s ± 0.0076s (3.5% slower)
  • Speedup: 0.9661 (< 1.0, threading hurt)
  • Contention: < 5%

PyTorch's internal thread pool already handles parallelism adequately. CPU-bound PyTorch ops (matmul, FFT) benefit from OpenMP/MKL threading, not from OS-level affinity. The <5% contention confirms no significant scheduler overhead.

Consequences

  • diffcfd/utils/threading.py retained as utility (no removal needed)
  • No thread affinity in production code
  • If future profiling at larger grid sizes (256²+) shows contention, reopen with fresh data

Related

  • scripts/profile_results.json
  • scripts/profile_thread_contention.py