Why is Python generally slower than compiled languages like C++ for CPU-bound tasks, and how can this be mitigated?
Interpretation, dynamic typing, and boxing add per-operation overhead; mitigate with C extensions, vectorization, or compilation
CPython compiles Python source to bytecode, not machine code, and executes that bytecode in a loop written in C. Each bytecode is a dynamic dispatch: the interpreter must look up the operation on a type object, check argument types, and allocate a new result object. Numbers are heap-allocated PyLong and PyFloat objects, so a simple addition is a function call plus an allocation rather than a machine add. Attribute lookups go through the MRO and the descriptor protocol. None of these individual costs is large, but they apply per operation, and a tight loop executes millions of them. C++ pays those costs once at compile time and emits direct machine instructions with registers and stack values. That is the whole reason for the gap: not that Python is poorly implemented, but that it trades runtime dynamism for compile-time specialization.
- 1
Move the hot loop out of Python: use NumPy, Cython, or a C extension so the loop runs in compiled code.
- 2
Use built-ins and stdlib functions where possible. They are implemented in C and avoid per-item bytecode.
- 3
Reduce allocation: preallocate lists, avoid building intermediate objects, and reuse buffers.
- 4
Cache attribute lookups and method references in hot paths, since lookup is not free.
- 5
Consider PyPy for long-running pure Python CPU work, or Numba for numeric kernels, but verify library compatibility.
- 6
Trade-off: C extensions reduce portability and add build complexity. Vectorization requires the problem to fit an array model.
- 7
Common mistake: micro-optimizing Python syntax while leaving an O(n^2) algorithm in place. Fix the algorithm first.
- 8
Version note: Python 3.11 introduced an adaptive specializing interpreter and 3.13 continued this work, narrowing the gap for some workloads but not eliminating it.
Scenario Questions
0-2 years experience
- 1Why is a Python for loop over a million integers slower than the equivalent C loop?
- 2What does it mean that Python integers are heap-allocated objects?
2-5 years experience
- 1You optimize a Python function by rewriting the syntax but see no improvement. What do you check first?
- 2You have a numeric inner loop that dominates runtime. What are your options beyond rewriting it in Python?
5-8 years experience
- 1You need to accelerate a kernel without leaving the Python ecosystem. How do you choose between NumPy, Cython, Numba, and a C extension?
- 2You are packaging a C extension for multiple platforms. What build and compatibility concerns arise?
8+ years experience
- 1Design a performance strategy for a CPU-heavy Python service that balances developer velocity, portability, and runtime speed.
- 2Explain how CPython's specializing interpreter and free-threaded builds change the classic performance picture, and how to validate the benefit for your workload.
Follow-up Questions
- When would PyPy help and when would it hurt compared to CPython?
- How do you decide whether to write a C extension versus using Cython or Numba?