fdiblen Performance profiling and optimization
The cardinal rule: never optimize unprofiled code. Intuition about where research code spends time is wrong often enough that acting on it wastes effort and adds complexity to the wrong places. The discipline is a loop: measure, find the dominant cost, fix only that, measure again, stop when it is fast enough for the science. "Fast enough" is a research judgment - a one-off analysis needs no tuning; a per-particle inner loop run for months justifies serious work.
Profile first
- Whole-program view: a sampling profiler (py-spy for Python:
py-spy top --pidon a live process orpy-spy recordfor a flame graph) shows where time actually goes with negligible overhead and no code changes. Language-native equivalents: perf (compiled code), Rprof (R), @profile/BenchmarkTools (Julia). - Line/function view once a hotspot is known: cProfile+snakeviz, line-profiler, or the IDE's profiler.
- Memory matters separately: allocation churn and swapping look like CPU slowness; profile memory when RSS grows or the machine swaps.
- Profile REPRESENTATIVE inputs at meaningful scale - toy inputs have different hotspots than production ones.
Report findings as fractions ("68% of runtime in distance_matrix"), not feelings; the fraction bounds the possible speedup (Amdahl's law) and justifies - or kills - the optimization.
Optimize in payoff order
- Algorithmic: a better algorithm or data structure beats any micro-optimization; check complexity before code details.
- Do less: cache repeated computation, hoist work out of loops, short-circuit, avoid recomputing unchanged pipeline stages (rseng-workflows).
- Use optimized building blocks: vectorized NumPy/BLAS operations, compiled library routines - most scientific speedups come from replacing interpreted loops with library calls (rseng-language-guides).
- Compile the hotspot: Numba/Cython/C extension for the proven hot function only.
- Parallelize last, and match the tool to the shape: threads/ processes on one node, job arrays or MPI on clusters (rseng-hpc-computing), GPUs for data-parallel numeric work (rseng-gpu-computing), distributed frameworks for larger-than- memory data (rseng-big-data-processing).
Preserve correctness at each step: run the tests (rseng-testing) after every optimization, with numerical tolerances where results legitimately differ (bit-identical is often the wrong bar - rseng-numerical-accuracy).
Keep it fast: benchmark tracking
One-off optimization decays; regressions arrive silently in innocent-looking commits. airspeed velocity (asv) runs a benchmark suite across commits, tracks results over time and flags regressions - the performance analogue of a test suite. Start small: benchmark the 3-5 operations users actually wait on, wire the run into CI or a scheduled job (rseng-ci-cd), and treat an unexplained regression like a failing test.
For quick comparisons during work, timeit/hyperfine with repeated runs and reported variance beat single stopwatch numbers; pin the environment (rseng-reproducible-environments) so timings compare across machines honestly - and record the hardware in any published benchmark.
Scaling measurements
Before requesting bigger allocations, measure scaling: run the same problem at 1, 2, 4, 8 workers (strong scaling) or grow the problem with workers (weak scaling), and plot efficiency. Recommending more hardware without a scaling curve is guesswork; the curve also feeds honest resource requests (rseng-hpc-computing) and energy accounting (rseng-green-computing).
Working with this skill
This skill is source-independent: its authority is the profiler and benchmark-tool documentation linked below and standard performance engineering practice.
Learn more (verified):
- https://github.com/benfred/py-spy - sampling profiler for Python
- https://github.com/airspeed-velocity/asv - benchmark regression tracking
- https://github.com/sharkdp/hyperfine - hyperfine command-line benchmarking
Related skills
Check whether any of these applies before moving on:
- rseng-big-data-processing - memory-bound escalation path
- rseng-ci-cd - benchmark regression tracking in CI
- rseng-gpu-computing - GPU port only after profiling evidence
- rseng-green-computing - speedups cut energy proportionally
- rseng-hpc-computing - scaling curves before big allocations
- rseng-numerical-accuracy - tolerances when optimizations shift results