static_vector vs vector<unique_ptr> Benchmark

You can read more about static_vector here.

A tiny benchmark suite that compares the performance of a hand‑rolled static_vector<T> (a fixed‑capacity, contiguous container) against a standard std::vector<std::unique_ptr<T>> when storing a non‑copyable, non‑movable Dummy type.

The benchmark measures three common patterns:

Metric	Description
Create + Destroy	Time to allocate, construct every element, then destruct and deallocate the container.
Iterate	Sequentially traverse the container with a range‑based for loop (read‑only).
Random Access	Access elements through an index array filled with uniform random indices.

For each metric the program reports the mean and a 95 % confidence interval (CI) across configurable trial runs.

---

Requirements

Tool	Version & Notes
C++ compiler	C++23 support is required for <print> and deducing‑this function syntax. ‒ GCC 14+ or Clang 18+ are known to work.

---

Building the Benchmark

$ g++ -std=c++23 -O3 -march=native -o benchmark main.cpp

Feel free to swap in clang++ or adjust optimisation flags for your platform.

---

Running

Human‑readable table

$ ./benchmark

CSV output (for plotting)

$ ./benchmark --csv > results.csv

The program benchmarks the following container sizes (elements):

1, 10, 100, 1 000, 10 000, 100 000, 1 000 000, 10 000 000, 100 000 000

Each measurement is repeated 100 times by default; edit trials in main.cpp if you need a different sample size.

Note  The helper clean_cache() function allocates ≈17 MiB and scribbles pseudo‑random data between runs to reduce cache‑warm effects. Make sure your system does not have more than 17MiB cache.

---

Plotting Results

$ python plot.py --filename results.csv

The script produces benchmark_blog_plot.png—a three‑panel log‑log plot with shaded 95 % CIs. On my system it looks like this: