From d1c28c2f4ebf1ba6a2084f0a1a44011daeb42277 Mon Sep 17 00:00:00 2001
From: oech3 <79379754+oech3@users.noreply.github.com>
Date: Thu, 30 Apr 2026 22:37:40 +0900
Subject: [PATCH] yes: add BENCHMARKING.md

---
 src/uu/yes/BENCHMARKING.md | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)
 create mode 100644 src/uu/yes/BENCHMARKING.md

diff --git a/src/uu/yes/BENCHMARKING.md b/src/uu/yes/BENCHMARKING.md
new file mode 100644
index 0000000000..d8fa624865
--- /dev/null
+++ b/src/uu/yes/BENCHMARKING.md
@@ -0,0 +1,24 @@
+# Benchmarking yes
+
+`yes` is a utility printing the provided string and a newline continuously.
+The default value `y` is used to skip the common `y/n` prompts provided by scripts.
+
+## Understanding yes
+
+The most simple way to implement `yes` is the `println!` loop. But it is slow since it
+calls many write syscalls. So `yes` should print an extended string to few bytes in the loop for
+better throughput.
+
+It is difficult to use `hyperfine` for benchmarking `yes` as it has infinite loop.
+But you can see throughput of `yes` by `pv` instead:
+
+```shell
+yes | pv >/dev/null
+```
+
+### `tee()` zero-copy
+
+The `read()` and `write()` based `yes` implementation is much slower than RAM's bandwidth
+since those syscalls are copying content of RAM.
+
+On Linux, `tee()` and `splice()` for non-pipe output is used to avoid copying content of RAM.