ci / gate (push) Waiting to run
- 项目与文档标题:nax-rs → ssw-mpc(尚数网-多方安全计算),二进制 nax → mpc - 18 个 crate 及目录 nax-* → mpc-*,Rust 引用 nax_* → mpc_* - 环境变量、C 符号、构建/部署脚本:NAX_* → MPC_*、libnax_ray.so → libmpc_ray.so - 保留协议域分隔常量与 refs/nax 参考实现路径,维持 golden 向量与历史数据兼容 - 新增 [profile.dev] incremental=false
157 lines
5.3 KiB
Rust
157 lines
5.3 KiB
Rust
//! 性能基准 —— **只打印,不断言**。
|
||
//!
|
||
//! # 为什么要单独一个文件
|
||
//!
|
||
//! `docs/` 和 memory 里那张「Python vs Rust」对照表,之前是临时手测出来贴上去的。
|
||
//! 那种数字有两个问题:**别人复现不了**,以及**代码改了没人知道它退化了**。
|
||
//! 这个文件把那几只数字变成一条命令:
|
||
//!
|
||
//! ```text
|
||
//! cargo test -p mpc-core --release --test bench_core -- --ignored --nocapture
|
||
//! ```
|
||
//!
|
||
//! # 为什么 `#[ignore]`
|
||
//!
|
||
//! 基准在共享 CI runner 上的数字没有意义(邻居抢 CPU、容器限频),
|
||
//! 而且会让 `cargo test` 从「秒级」变成「几十秒」。默认不跑,要跑显式点名。
|
||
//!
|
||
//! # 读数字的注意事项
|
||
//!
|
||
//! * 必须 `--release`。debug 下的环乘法慢一个数量级,数字纯属误导 ——
|
||
//! 所以本文件开头会检查 `debug_assertions` 并直接提示。
|
||
//! * `matmul` 在 x86_64 上走 AVX2(运行时探测),armv7 走标量。
|
||
//! 两边的数字不可比,报告里要写清楚测的是哪台机。
|
||
|
||
use std::time::Instant;
|
||
|
||
use mpc_core::aead;
|
||
use mpc_core::prg::Stream;
|
||
use mpc_core::ring;
|
||
use mpc_core::share;
|
||
|
||
/// 跑 `reps` 次取最好成绩(排除首次的缺页/预热)。
|
||
fn best<F: FnMut()>(reps: usize, mut f: F) -> f64 {
|
||
let mut best_ms = f64::MAX;
|
||
for _ in 0..reps {
|
||
let t = Instant::now();
|
||
f();
|
||
let ms = t.elapsed().as_secs_f64() * 1e3;
|
||
if ms < best_ms {
|
||
best_ms = ms;
|
||
}
|
||
}
|
||
best_ms
|
||
}
|
||
|
||
fn row(name: &str, scale: &str, ms: f64, wl: &str) {
|
||
println!("{name:<26} {scale:>14} {ms:>9.3} ms {wl}");
|
||
}
|
||
|
||
#[test]
|
||
#[ignore = "性能基准:需要 --release 且应独占 CPU"]
|
||
fn bench() {
|
||
if cfg!(debug_assertions) {
|
||
println!("\n⚠ 当前是 debug 构建,环乘法的数字会比 release 慢一个数量级。");
|
||
println!(" 正确跑法:cargo test -p mpc-core --release --test bench_core -- --ignored --nocapture\n");
|
||
}
|
||
|
||
let mut s = Stream::new(&[7u8; 32]);
|
||
|
||
// ---- 环算术:单元素 ----
|
||
let n = 1 << 20; // 1M 元素
|
||
let a: Vec<u64> = (0..n).map(|i| (i as u64).wrapping_mul(0x9E37_79B9_7F4A_7C15)).collect();
|
||
let b: Vec<u64> = (0..n).map(|i| (i as u64).wrapping_mul(0xC2B2_AE3D_27D4_EB4F)).collect();
|
||
let mut out = vec![0u64; n];
|
||
|
||
let ms = best(3, || ring::mul_into(&a, &b, &mut out).unwrap());
|
||
row("逐元素 mul", "1M u64", ms, &format!("{:.2} G elem/s", n as f64 / ms / 1e6));
|
||
|
||
let ms = best(3, || ring::add_into(&a, &b, &mut out).unwrap());
|
||
row("逐元素 add", "1M u64", ms, &format!("{:.2} G elem/s", n as f64 / ms / 1e6));
|
||
|
||
// ---- 矩阵乘 ----
|
||
for side in [256usize, 512] {
|
||
let m = side * side;
|
||
let k = side * side;
|
||
let am: Vec<u64> = (0..m as u64).collect();
|
||
let bm: Vec<u64> = (0..k as u64).collect();
|
||
let mut om = vec![0u64; m];
|
||
let reps = if side >= 512 { 2 } else { 5 };
|
||
let ms = best(reps, || {
|
||
ring::matmul_into(side, side, side, &am, &bm, &mut om).unwrap()
|
||
});
|
||
let gop = (side as f64).powi(3) / ms / 1e6;
|
||
row(
|
||
"matmul",
|
||
&format!("{side}³"),
|
||
ms,
|
||
&format!("{gop:.2} G mul-add/s"),
|
||
);
|
||
}
|
||
|
||
// ---- 分片 / 重建 ----
|
||
let secret: Vec<u64> = (0..n as u64).collect();
|
||
let mut shares = vec![0u64; 2 * n];
|
||
let ms = best(3, || {
|
||
share::split_into(&secret, 2, &mut s, &mut shares, 18).unwrap();
|
||
});
|
||
row("split_into", "1M u64 / 2 方", ms, &format!("{:.2} G elem/s", n as f64 / ms / 1e6));
|
||
|
||
let ms = best(3, || {
|
||
share::reconstruct_into(&shares, 2, n, &mut out).unwrap();
|
||
});
|
||
row("reconstruct_into", "1M u64 / 2 方", ms, &format!("{:.2} G elem/s", n as f64 / ms / 1e6));
|
||
|
||
// ---- ChaCha20 密钥流(MPC 随机性的底座)----
|
||
let mut buf = vec![0u8; 16 << 20];
|
||
let ms = best(3, || {
|
||
let mut st = Stream::new(&[3u8; 32]);
|
||
st.fill_bytes(&mut buf);
|
||
});
|
||
row(
|
||
"ChaCha20 流",
|
||
"16 MiB",
|
||
ms,
|
||
&format!("{:.0} MiB/s", 16.0 / (ms / 1e3)),
|
||
);
|
||
|
||
// ---- AEAD:大块吞吐 + 小帧延迟 ----
|
||
let key = [9u8; 32];
|
||
let nonce = [1u8; 12];
|
||
let plain = vec![0x5Au8; 16 << 20];
|
||
let mut sealed = vec![0u8; plain.len() + aead::TAG_LEN];
|
||
let ms = best(3, || {
|
||
aead::seal(&key, &nonce, b"", &plain, &mut sealed).unwrap();
|
||
});
|
||
row(
|
||
"AEAD seal",
|
||
"16 MiB",
|
||
ms,
|
||
&format!("{:.0} MiB/s", 16.0 / (ms / 1e3)),
|
||
);
|
||
// 输出缓冲在循环外分配:每次 open 都分配 16MB 的话,测到的是分配器不是密码学
|
||
let mut open_out = vec![0u8; plain.len()];
|
||
let ms = best(3, || {
|
||
aead::open(&key, &nonce, b"", &sealed, &mut open_out).unwrap();
|
||
});
|
||
row(
|
||
"AEAD open",
|
||
"16 MiB",
|
||
ms,
|
||
&format!("{:.0} MiB/s", 16.0 / (ms / 1e3)),
|
||
);
|
||
|
||
// 小帧是 MPC 的常态(一帧通常几十字节到几十 KB)
|
||
let small = vec![0u8; 1024];
|
||
let mut small_out = vec![0u8; small.len() + aead::TAG_LEN];
|
||
let reps = 2000;
|
||
let t = Instant::now();
|
||
for _ in 0..reps {
|
||
aead::seal(&key, &nonce, b"", &small, &mut small_out).unwrap();
|
||
}
|
||
let us = t.elapsed().as_secs_f64() * 1e6 / reps as f64;
|
||
row("AEAD seal", "1 KiB × 2000", us / 1e3, &format!("{us:.2} µs/帧"));
|
||
|
||
println!();
|
||
}
|