为何Numba指定parallel=True后range仍未并行化，需用prange？

Great question—let’s unpack exactly what’s going on with your test code and why only f4 is leveraging multiple CPU cores effectively.

Let’s start by analyzing each function’s behavior: #

• f1: This is a standard single-threaded Numba-compiled function. No parallelism is enabled, so it uses only one CPU core (100% usage) and takes the longest to run.
• f2: You used prange but forgot to add parallel=True to the njit decorator. Without this flag, prange is just an alias for regular range—no parallel compilation happens, hence the runtime is nearly identical to f1.
• f3: You enabled parallel=True but stuck with regular range loops. Numba’s automatic parallelization is intentionally conservative: it won’t automatically parallelize nested range loops unless it can definitively confirm there are no iteration dependencies. For complex nested structures like yours, it doesn’t make that call, so it stays single-threaded.
• f4: Here you combined parallel=True with prange for all loops. prange acts as an explicit signal to Numba: “This loop’s iterations are independent and safe to run in parallel.” With this marker, Numba splits the work across multiple cores, leading to ~300% CPU usage and the fastest runtime.

Why range doesn’t auto-parallelize nested loops with parallel=True: #

The official docs examples that show range working with parallel=True are almost always simple, single-level loops where Numba can easily detect independence. For nested loops:

1. Dependency ambiguity: Numba needs to guarantee that no iteration relies on the result of another. While your calculation is technically independent per iteration, Numba’s auto-parallelizer doesn’t reliably pick up on this for nested range loops.
2. Explicit control: prange gives you control over which loops to parallelize. Parallelizing inner loops often has higher overhead than outer ones, so Numba leaves the decision to you via prange instead of guessing.

Quick rules of thumb for Numba parallelism: #

• Always pair prange with parallel=True—one without the other won’t enable parallelism.
• Use prange for outer loops first (they offer the most benefit for parallelization).
• parallel=True alone only enables parallel compilation mode; it doesn’t force Numba to parallelize every loop.

内容的提问来源于stack exchange，提问作者Labo