Stack vs malloc: real-world benchmark shows 2–6x difference

Usually, we assume that malloc is fast—and in most cases it is.
However, sometimes "reasonable" code can lead to very unreasonable performance.

In a previous post, I looked at using stack-based allocation (VLA / fixed-size) for temporary data, and another on estimating available stack space to use it safely.

This time I wanted to measure the actual impact in a realistic workload.

Full Article (Medium - no paywall):

Stack vs malloc: real-world benchmark shows 2–6x difference

I built a benchmark based on a loan portfolio PV calculation, where each loan creates several temporary arrays (thousands of elements each). This is fairly typical code-clean, modular, nothing unusual.

I compared:

• stack allocation (VLA)
• heap per-loan (malloc/free)
• heap reuse
• static (baseline)

Results:

• stack allocation stays very close to optimal
• heap per-loan can be ~2.5x slower (glibc) and up to ~6x slower (musl)
• even optimized allocators show pattern-dependent behavior

The main takeaway for me: allocation cost is usually hidden—but once it's in the hot path, it really matters.

Curious how others approach temporary workspace in performance-sensitive code.