Skip to content

PROFILING.md

Latest commit

 

History

History
92 lines (65 loc) · 3.05 KB

PROFILING.md

File metadata and controls

92 lines (65 loc) · 3.05 KB
Raw

Profiling

How to profile QCTidy with samply, a sampling profiler that shows a flame graph in the browser.

Install samply

cargo install samply

On Linux, samply also needs kernel.perf_event_paranoid set to 1 or lower. If profiling fails with a perf_event_paranoid message, lower it with:

echo '1' | sudo tee /proc/sys/kernel/perf_event_paranoid

This change is temporary (it resets on the next reboot) and is a system-wide security tweak; restore it to 2 when you are done.

Build a release binary with symbols

Profiling works best on an optimized build that still keeps function names and line numbers. Set CARGO_PROFILE_RELEASE_DEBUG for the line info and override the project's strip = true with CARGO_PROFILE_RELEASE_STRIP=none (otherwise the binary is stripped and samply only sees raw addresses):

CARGO_PROFILE_RELEASE_DEBUG=1 CARGO_PROFILE_RELEASE_STRIP=none cargo build --release -p qctidy-cli

The strip = true comes from [profile.release] in the root Cargo.toml, so it must be overridden for profiling.

Generate a reproducible input

A tiny circuit finishes too fast to profile. Generate a large, deterministic Qiskit circuit instead:

cargo run --release -p benchmarks --bin generate_circuit -- -o /tmp/profile_circuit.py

The defaults (50 qubits, 1000 time steps, seed 42) produce a circuit of about 34 000 lines that qctidy check takes roughly 5 minutes to process. The output only depends on the seed, so the circuit is reproducible.

Tune the size and seed with the flags:

cargo run --release -p benchmarks --bin generate_circuit -- \
  --seed 7 --qubits 100 --time-steps 500 -o /tmp/profile_circuit.py

Options: --seed, --qubits, --time-steps, --circuit-name, and -o/--output (writes to standard output when -o is omitted).

The generator builds gate-only circuits (no measurements), because the fixer — the usual hot path — does not touch measurements.

Profile a command

Record the profile to a file, then load it with samply load, which serves the symbols:

samply record --save-only -o check.json -- target/release/qctidy check -i /tmp/profile_circuit.py
samply load check.json

samply load opens the browser and supplies symbols by reading the binary's debug info. Loading check.json directly in https://profiler.firefox.com/ (drag and drop) does not show symbols: the online profiler cannot read the local binary, so it only shows raw addresses.

Other commands worth profiling:

# Metrics (analyzer)
samply record target/release/qctidy analyze -i /tmp/profile_circuit.py --detailed

# Parse/serialize (converter)
samply record target/release/qctidy convert -i /tmp/profile_circuit.py --output-format json

Reading the flame graph

The widest bars are the functions that spend the most CPU time. check on the generated circuit scales worse than linearly with the number of gates, so the bottleneck is somewhere in the fixer's pattern matching. Look for the widest frames under fixer and matcher.