perf(hotpath): add cpu_hotspot_ring bench and send-chain optimization plan

- Add measure_all to PeerMap and CidrSet impl blocks (hotpath::measure_all)
- Add [profile.hotpath] for samply-compatible builds (strip=false, debug=line-tables-only)
- Add cpu_hotspot_ring example: 2-node ring tunnel with data-plane flooding (~234K pps)
- Add plans/006-send-chain-cpu-optimization.md based on hotpath+samply 423M sample analysis
  Key findings: dashmap redundancy (14.9%), metrics overhead (8.3%), mpsc (14.1%)
  Target: reduce send_msg_internal from 3.26us to ~2us per packet
This commit is contained in:
fanyang
2026-06-28 12:30:50 +08:00
parent 7205517160
commit 79035ea972
5 changed files with 363 additions and 2 deletions
+6
View File
@@ -27,3 +27,9 @@ lto = true
codegen-units = 1
opt-level = 3
strip = true
# For hotpath CPU profiling: samply needs debug symbols and unstripped binaries.
[profile.hotpath]
inherits = "release"
strip = false
debug = "line-tables-only"
+155
View File
@@ -0,0 +1,155 @@
//! CPU hotspot benchmark for hotpath profiling.
//!
//! Builds two no-tun EasyTier instances connected via an in-process ring
//! tunnel, lets routes converge, then floods data-plane packets through
//! `send_msg_by_ip` so that `hotpath-cpu` / samply can collect meaningful
//! CPU samples.
//!
//! Build & run:
//! cargo run --profile hotpath --features hotpath,hotpath-cpu \
//! --example cpu_hotspot_ring
//!
//! Prerequisites: hotpath-samply + samply must be installed and on PATH.
//! See bench/006-hotpath-cpu-top.md for install instructions.
//!
//! Then in another terminal:
//! hotpath console
use std::net::IpAddr;
use std::time::{Duration, Instant};
use bytes::BytesMut;
use easytier::common::config::{ConfigLoader, TomlConfigLoader};
use easytier::instance::instance::Instance;
use easytier::tunnel::packet_def::ZCPacket;
use easytier::tunnel::ring::RingTunnelConnector;
#[tokio::main(flavor = "multi_thread", worker_threads = 4)]
#[cfg_attr(feature = "hotpath", hotpath::main)]
async fn main() {
let duration = std::env::var("HOTPATH_BENCH_SECS")
.ok()
.and_then(|s| s.parse().ok())
.unwrap_or(30u64);
let pkt_size: usize = std::env::var("HOTPATH_PKT_SIZE")
.ok()
.and_then(|s| s.parse().ok())
.unwrap_or(1400);
let mut inst_a = Instance::new(no_tun_config("hot-a", "10.144.144.1"));
let mut inst_b = Instance::new(no_tun_config("hot-b", "10.144.144.2"));
inst_a.run().await.expect("inst_a run");
inst_b.run().await.expect("inst_b run");
let ring_url = format!("ring://{}", inst_a.id());
inst_b
.get_conn_manager()
.add_connector(RingTunnelConnector::new(ring_url.parse().unwrap()));
let dst: IpAddr = "10.144.144.2".parse().unwrap();
let src = "10.144.144.1";
let converged = tokio::time::timeout(Duration::from_secs(15), async {
loop {
let a = inst_a.get_peer_manager().list_routes().await;
let b = inst_b.get_peer_manager().list_routes().await;
if a.len() >= 1 && b.len() >= 1 {
return true;
}
tokio::time::sleep(Duration::from_millis(500)).await;
}
})
.await
.is_ok();
if !converged {
eprintln!("warning: routes did not converge within 15s");
}
println!(
"cpu_hotspot_ring: flooding {}s, pkt_size={} (converged={})",
duration, pkt_size, converged
);
let pm = inst_a.get_peer_manager();
let send_pkt = make_data_packet(src, "10.144.144.2", pkt_size);
let sender_task = tokio::spawn(async move {
let mut sent: u64 = 0;
let start = Instant::now();
loop {
let pkt = send_pkt.clone();
let _ = pm.send_msg_by_ip(pkt, dst, false).await;
sent += 1;
if sent % 10000 == 0 {
let elapsed = start.elapsed().as_secs_f64();
let pps = sent as f64 / elapsed;
let mbps = pps * pkt_size as f64 * 8.0 / 1_000_000.0;
println!("sent {} pkts ({:.0} pps, {:.0} Mbps)", sent, pps, mbps);
}
}
});
tokio::time::sleep(Duration::from_secs(duration)).await;
sender_task.abort();
println!("cpu_hotspot_ring: done");
}
fn make_data_packet(src: &str, dst: &str, total_size: usize) -> ZCPacket {
use std::net::Ipv4Addr;
let hdr_len = 28;
let payload_len = total_size.saturating_sub(hdr_len);
let ip_total_len = (hdr_len + payload_len) as u16;
let mut buf = BytesMut::with_capacity(total_size);
buf.extend_from_slice(&[
0x45,
0x00,
(ip_total_len >> 8) as u8,
(ip_total_len & 0xff) as u8,
0x00,
0x00,
0x40,
0x00,
0x40,
0x11,
0x00,
0x00,
]);
let src: Ipv4Addr = src.parse().unwrap();
buf.extend_from_slice(&src.octets());
let dst: Ipv4Addr = dst.parse().unwrap();
buf.extend_from_slice(&dst.octets());
let udp_len = (8 + payload_len) as u16;
buf.extend_from_slice(&[
0x30,
0x39,
0xD4,
0x31,
(udp_len >> 8) as u8,
(udp_len & 0xff) as u8,
0x00,
0x00,
]);
buf.resize(total_size, 0xAA);
ZCPacket::new_with_payload(&buf)
}
fn no_tun_config(name: &str, ipv4: &str) -> TomlConfigLoader {
let config = TomlConfigLoader::default();
config.set_inst_name(name.to_owned());
config.set_ipv4(Some(ipv4.parse().unwrap()));
let mut flags = config.get_flags();
flags.no_tun = true;
config.set_flags(flags);
config
}
+1
View File
@@ -32,6 +32,7 @@ pub(crate) struct CidrSet {
mapped_to_real: Arc<DashMap<cidr::Ipv4Cidr, cidr::Ipv4Cidr>>,
}
#[cfg_attr(feature = "hotpath", hotpath::measure_all)]
impl CidrSet {
pub fn new(global_ctx: ArcGlobalCtx) -> Self {
let mut ret = Self {
+1 -2
View File
@@ -38,6 +38,7 @@ pub struct PeerMap {
alive_client_urls: Arc<Mutex<multimap::MultiMap<url::Url, PeerConnId>>>,
}
#[cfg_attr(feature = "hotpath", hotpath::measure_all)]
impl PeerMap {
pub fn new(packet_send: PacketRecvChan, global_ctx: ArcGlobalCtx, my_peer_id: PeerId) -> Self {
PeerMap {
@@ -132,7 +133,6 @@ impl PeerMap {
peer_id == self.my_peer_id || self.peer_map.contains_key(&peer_id)
}
#[cfg_attr(feature = "hotpath", hotpath::measure(impl_type = "PeerMap"))]
pub async fn send_msg_directly(&self, msg: ZCPacket, dst_peer_id: PeerId) -> Result<(), Error> {
if dst_peer_id == self.my_peer_id {
let packet_send = self.packet_send.clone();
@@ -164,7 +164,6 @@ impl PeerMap {
Ok(())
}
#[cfg_attr(feature = "hotpath", hotpath::measure(impl_type = "PeerMap"))]
pub async fn get_gateway_peer_id(
&self,
dst_peer_id: PeerId,
+200
View File
@@ -0,0 +1,200 @@
# 计划 006send_msg_internal 发包链路 CPU 优化
> **执行者说明**:按步骤执行本计划。每一步都必须运行验证命令,并确认结果符合预期后再继续。完成后更新 `plans/README.md` 中本计划的状态行。
## 状态
- **优先级**: P1
- **工作量**: M
- **风险**: LOW
- **依赖**: hotpath profiling infra (main branch commit `be2034dd`)
- **类别**: performance
- **数据来源**: hotpath-cpu samply 423,583,601 samples4 threads234K ppspkt_size=1400
## 为什么重要
`send_msg_internal` 是数据面包转发的核心路径,每包耗时 3.26µs(wall time)。在 234K pps 下占 wall time 的 ~70%。samply inclusive CPU 分解显示有多处可通过减少冗余操作来省 µs 级开销。每包省 1µs 即可将吞吐提升 ~30%。
## 数据基线
### timingwall time,含 await
| Function | Calls | Avg/包 | 级差 |
|---|---|---|---|
| `send_msg_internal` | 6.9M | 3.26µs | — |
| └─ `send_msg_directly` | 6.9M | 2.83µs | 0.43µs(路由决策) |
| └─ `Peer::send_msg` | 6.9M | 2.69µs | 0.14µsconn 选择) |
| └─ `PeerConn::send_msg` | 6.9M | 2.58µs | 0.11µssession 选择) |
### samply inclusive CPUsend_msg_internal 子树,11.5M samples
| % | Function | 含义 |
|---|---|---|
| 12.0% | `PeerMap::send_msg_directly` | 发包核心 |
| 7.5% | `tokio::mpsc::Sender::send` | mpsc 通道 |
| **7.1%** | **`TrafficMetricRecorder::record_tx`** | 每包流量统计 |
| **6.1%+4.8%+4.0%** | **`dashmap::get` ×3** | 冗余 dashmap 查询 |
| 5.6% | `batch_semaphore::Acquire::poll` | mpsc permit |
| **3.9%** | **`quanta::get_now`** | 时间戳获取 |
| 3.9% | `malloc` | 内存分配 |
| **1.2%** | **`TrafficCounters closure`** | 流量计数器 |
| 1.0% | `MpscTunnelSender::send` | tunnel 发送 |
## 当前代码
```rust
// easytier/src/peers/peer_manager.rs:1533-1588
async fn send_msg_internal(
peers: &Arc<PeerMap>,
foreign_network_client: &Arc<ForeignNetworkClient>,
relay_peer_map: &Arc<RelayPeerMap>,
direct_tx_metrics: Option<&Arc<TrafficMetricRecorder>>,
msg: ZCPacket,
dst_peer_id: PeerId,
) -> Result<(), Error> {
// ...
let send_result = if ... {
// relay path
} else if peers.has_peer(dst_peer_id) { // dashmap get #1 (contains_key)
peers.send_msg_directly(msg, dst_peer_id).await // 内部 get_peer_by_id = dashmap get #2
} else if foreign_network_client.has_next_hop(dst_peer_id) {
// foreign network path
} else if let Some(gateway) = peers.get_gateway_peer_id(dst_peer_id, policy.clone()).await {
if peers.has_peer(gateway) || ... { // dashmap get #3
relay_peer_map.send_msg(msg, dst_peer_id, policy).await
}
}
if send_result.is_ok() && let Some(metrics) = direct_tx_metrics {
metrics.record_tx(dst_peer_id, packet_type, msg_len).await; // 每包记录
}
send_result
}
```
```rust
// easytier/src/peers/peer_map.rs:136-164
pub async fn send_msg_directly(&self, msg: ZCPacket, dst_peer_id: PeerId) -> Result<(), Error> {
if dst_peer_id == self.my_peer_id {
// self-send path (tokio::spawn)
return Ok(());
}
match self.get_peer_by_id(dst_peer_id) { // dashmap get (重复)
Some(peer) => peer.send_msg(msg).await?,
None => return Err(Error::RouteError(...)),
}
Ok(())
}
```
## 优化项
### 步骤 1:合并 dashmap 冗余查询(P0,预期省 ~0.1-0.2µs/包)
**问题**happy path 上 `has_peer(dst_peer_id)` + `send_msg_directly → get_peer_by_id(dst_peer_id)` 对同一个 key 做了 2 次 dashmap 查询。每次 ~100nshash + shard read lock)。
**方案**:在 `send_msg_internal` 中直接调 `get_peer_by_id`,根据 `Option<Arc<Peer>>` 分支,跳过 `has_peer` 检查。
```rust
// 改前
} else if peers.has_peer(dst_peer_id) {
peers.send_msg_directly(msg, dst_peer_id).await
}
// 改后
} else if let Some(peer) = peers.get_peer_by_id(dst_peer_id) {
peer.send_msg(msg).await
}
```
注意:`send_msg_directly` 中的 self-send 分支(`dst_peer_id == my_peer_id`)需要在上层处理或保留。当前 bench 场景 `dst_peer_id != my_peer_id`,不触发 self-send。
**涉及文件**`easytier/src/peers/peer_manager.rs:1558-1559`
**冲突检查**advisor/001-002 改过此文件(队列背压 + metrics 连带),需 rebase 后确认行号。
**验证**`cargo test -p easytier -- send_msg_internal`
### 步骤 2TrafficMetricRecorder 降频记录(P1,预期省 ~0.25µs/包)
**问题**`record_tx` 每包都调用,占 inclusive CPU 的 7.1% + TrafficCounters 1.2% = 8.3%。内部做 histogram 记录(`hdrhistogram::record_n_inner`)和时间戳获取(`quanta::get_now`)。
**方案**:在 `TrafficMetricRecorder` 中引入 per-thread atomic 计数器,每 N 包(如 64)或每 T ms 刷入 histogram。
```rust
// 改前
metrics.record_tx(dst_peer_id, packet_type, msg_len).await;
// 改后
metrics.record_tx_fast(dst_peer_id, packet_type, msg_len); // sync, atomic counter
// 内部: counter.fetch_add(msg_len); if counter % 64 == 0 { flush_to_histogram() }
```
**涉及文件**`easytier/src/peers/traffic_metrics.rs``easytier/src/peers/peer_manager.rs:1584`
**冲突检查**traffic_metrics.rs 零冲突。peer_manager.rs 同步骤 1。
**验证**`cargo test -p easytier -- traffic_metrics`
### 步骤 3:缓存时间戳(P2,预期省 ~0.13µs/包)
**问题**`quanta::get_now` 占 inclusive CPU 的 3.9%。send_msg_internal 路径上多处获取当前时间(record_tx 内部、traffic counters 等)。
**方案**:在 `send_msg_internal` 入口取一次时间戳,传入子函数。
```rust
let now = quanta::Instant::now();
// ...
metrics.record_tx_with_time(dst_peer_id, packet_type, msg_len, now);
```
**涉及文件**`easytier/src/peers/peer_manager.rs``easytier/src/peers/traffic_metrics.rs`
**冲突检查**:同步骤 2。
**验证**bench pps 对比。
### 步骤 4mpsc batch sendP3,预期省 ~0.46µs/包)
**问题**`PeerConn::send_msg` 每包做 1 次 `MpscTunnelSender::send`,触发 mpsc `Sender::send` (7.5%) + `batch_semaphore::Acquire::poll` (5.6%) + `add_permits_locked` (3.82%) = 16.9%。
**方案**:在 `PeerConn``Peer` 层引入 batch buffer,攒满 N 个包后一次 `send`(使用 `try_send` 或 unbounded channel)。
**涉及文件**`easytier/src/peers/peer_conn.rs``easytier/src/tunnel/mpsc.rs`
**冲突检查**peer_conn.rs 被 advisor/001-002 改过。mpsc.rs 被 perf/001 改过。需要协调合并顺序。
**验证**bench pps 对比 + `cargo test -p easytier -- peer_conn`
### 步骤 5ZCPacket 池化(P4,预期省 ~0.21µs/包)
**问题**:每包 malloc 3.9% + free 1.2% + morecore 1.2% = 6.3%。全局 munmap 4.73% 也部分来自此。
**方案**:对 ZCPacket 引入池化(`crossbeam-queue::ArrayQueue``tokio::sync::Pool`)。
**涉及文件**`easytier/src/tunnel/packet_def.rs`
**冲突检查**packet_def.rs 被 perf/001-003 改过。需要在 perf PR 合并后实施。
**验证**bench pps + `cargo test -p easytier -- packet`
## 预期总收益
| 步骤 | 每包省 | 累计 |
|---|---|---|
| 步骤 1dashmap 合并) | ~0.15µs | 3.26→3.11µs |
| 步骤 2metrics 降频) | ~0.25µs | 3.11→2.86µs |
| 步骤 3(缓存时间戳) | ~0.13µs | 2.86→2.73µs |
| 步骤 4batch send | ~0.46µs | 2.73→2.27µs |
| 步骤 5packet 池化) | ~0.21µs | 2.27→2.06µs |
| **合计** | **~1.2µs** | **3.26→2.06µs-37%** |
在 4 threads 配置下,预期 pps 从 234K 提升到 ~320K-370K+37%-58%)。
## 验证方法
```bash
# baseline(当前 main + measure_all
export PATH=$HOME/.cargo/bin:$PATH
cargo run --profile hotpath --features hotpath,hotpath-cpu --example cpu_hotspot_ring
# 记录 pps 和 timing avg
# 每个步骤实施后重跑,对比 pps 和 send_msg_internal avg
```
## 风险
- **步骤 1**:改变路由决策逻辑的边界条件(self-send、foreign network)。需确保不破坏 `send_msg_internal_*` 测试。
- **步骤 2**:metrics 精度降低(从每包精确变为每 64 包近似)。需确认 stats 查询端能接受。
- **步骤 4**batch send 引入延迟(攒批期间包等待)。需设置 flush timeout。
- **步骤 5**:ZCPacket 池化改变生命周期模型,可能引入 use-after-free。需充分测试。