perf-workshop/scenario1-python-to-c/README.md

Scenario 1: Python to C Optimization

Learning Objectives

• Use time to measure execution time
• Profile Python code with cProfile
• Generate flamegraphs with py-spy
• Use ctypes to call C code from Python

Files

• prime_slow.py - Pure Python implementation (slow)
• prime.c - C implementation of the hot function
• prime_fast.py - Python calling C via ctypes

Exercises

Step 1: Measure the baseline

time python3 prime_slow.py 100000

Note the real time (wall clock) and user time (CPU time in user space).

Step 2: Profile with cProfile

python3 -m cProfile -s cumtime prime_slow.py 100000

Look for:

• Which function has the highest cumtime (cumulative time)?
• How many times (ncalls) is is_prime called?

Step 3: Generate a flamegraph

py-spy record -o prime_slow.svg -- python3 prime_slow.py

Open prime_slow.svg in a browser. The widest bar at the top shows where time is spent.

Step 4: Compile and run the optimized version

# Compile the C library
gcc -O2 -fPIC -shared -o libprime.so prime.c

# Run the fast version
time python3 prime_fast.py 100000

Step 5: Compare

• How much faster is the C version?
• Generate a flamegraph for prime_fast.py - what's different?

Discussion Questions

1. Why is the C version faster? (Hint: interpreter overhead, type checking)
2. When is it worth rewriting in C vs. finding a library?
3. What's the trade-off of using ctypes vs native Python?