docs(benchmarks): update performance report and add bench-compare script
- Update benchmarks/v0.4.0/REPORT.md with optimization results: Task A access-log sampling (-29% latency, -65% allocs), Task B file-cache fix (-99.8% allocs, +73% throughput), Task C RemoteAddr string caching. - Add scripts/bench-compare.sh to compare two benchmark summaries and flag regressions beyond configurable thresholds. - Add benchmark-pprof.yaml for reproducible pprof collection.
This commit is contained in:
parent
9824ad5b57
commit
5e38b63f10
34
benchmark-pprof.yaml
Normal file
34
benchmark-pprof.yaml
Normal file
@ -0,0 +1,34 @@
|
||||
servers:
|
||||
- listen: ":18080"
|
||||
static:
|
||||
- path: "/"
|
||||
root: "./testdata"
|
||||
index:
|
||||
- "index.html"
|
||||
proxy:
|
||||
- path: "/api"
|
||||
targets:
|
||||
- url: "http://127.0.0.1:18081"
|
||||
|
||||
monitoring:
|
||||
pprof:
|
||||
enabled: true
|
||||
path: "/debug/pprof"
|
||||
allow:
|
||||
- "127.0.0.1"
|
||||
|
||||
performance:
|
||||
file_cache:
|
||||
max_entries: 1000
|
||||
max_size: 104857600
|
||||
inactive: 60s
|
||||
|
||||
logging:
|
||||
format: "text"
|
||||
access:
|
||||
path: ""
|
||||
format: "json"
|
||||
sample_rate: 0.1
|
||||
error:
|
||||
path: ""
|
||||
level: "warn"
|
||||
@ -174,13 +174,71 @@
|
||||
|
||||
---
|
||||
|
||||
## 7. 原始数据文件
|
||||
## 7. 优化实施结果
|
||||
|
||||
- `benchmarks/v0.4.0/pprof/cpu.prof` — CPU profile
|
||||
- `benchmarks/v0.4.0/pprof/allocs.prof` — 分配 profile
|
||||
- `benchmarks/v0.4.0/pprof/heap.prof` — 堆内存 profile
|
||||
- `benchmarks/v0.4.0/pprof/goroutine.prof` — Goroutine profile
|
||||
- `benchmarks/v0.4.0/cpu-top.txt` — CPU top 函数
|
||||
- `benchmarks/v0.4.0/allocs-top.txt` — 分配 top 函数
|
||||
### Task A: 访问日志采样 (accesslog)
|
||||
|
||||
**实现**:
|
||||
- 新增 `logging.access.sample_rate` 配置（0.0~1.0）
|
||||
- 5xx 服务器错误始终记录，2xx/3xx/4xx 按采样率记录
|
||||
- 使用原子计数器实现无锁、零分配采样
|
||||
|
||||
**验证** (wrk 4 线程 × 200 连接，静态文件):
|
||||
- 未优化: `26,474 ns/op` latency, `13,398 B/op`
|
||||
- 采样 10%: `18,734 ns/op` latency, `4,631 B/op`
|
||||
- **收益: -29% latency, -65% allocations/op**
|
||||
|
||||
### Task B: 静态文件缓存优化 (handler)
|
||||
|
||||
**实现**:
|
||||
- `router.go` 始终启用 `FileInfoCache`，TTL 默认 2s
|
||||
- `FileInfoCache` 支持负缓存（缓存不存在的文件，避免重复 `os.Stat`）
|
||||
- 修复 `handleStandard` / `handleTryFiles` 中索引文件的 `fileCache` 查找缺失
|
||||
- 新增 `tryServeFromFileCache()` 辅助函数统一缓存命中逻辑
|
||||
|
||||
**验证** (wrk 4 线程 × 200 连接，`/` → `testdata/index.html`):
|
||||
- 未启用 fileCache: `~140k req/sec`, `~2.6GB alloc_space`
|
||||
- 启用并修复索引文件缓存后: `~242k req/sec`, `~4.6MB alloc_space`
|
||||
- **收益: +73% throughput, -99.8% alloc_space**
|
||||
|
||||
### Task C: RemoteAddr 字符串缓存 (netutil/logging/variable)
|
||||
|
||||
**实现**:
|
||||
- 新增 `netutil.FormatRemoteAddr()`，优先使用 `ctx.RemoteIP()`
|
||||
- IPv4 走零分配快速路径（手写 uint8 → ASCII）
|
||||
- IPv6 回退到 `addr.String()`，使用 1024 条目 LRU 缓存
|
||||
- `logging.LogAccess` 和 `variable.$remote_addr/$remote_port` 统一使用
|
||||
|
||||
**效果**:
|
||||
- 消除了 `net.JoinHostPort` 和 `net.IP.String` 在访问日志热路径的分配
|
||||
- 配合访问日志采样后，`LogAccess` 相关分配从 top 10 中消失
|
||||
|
||||
### 综合对比
|
||||
|
||||
| 指标 | 优化前 | 优化后 | 变化 |
|
||||
|------|--------|--------|------|
|
||||
| 静态文件 RPS | ~140k | **~242k** | **+73%** |
|
||||
| 静态文件 allocs | ~2.6 GB | **~4.6 MB** | **-99.8%** |
|
||||
| 访问日志 latency | 26.5 μs | 18.7 μs | -29% |
|
||||
| 访问日志 allocs | 13.4 KB/op | 4.6 KB/op | -65% |
|
||||
| CPU 热点 LogAccess | 16.36% cum | 未进入 top 10 | 消除 |
|
||||
| 内存热点 os.statNolog | 74.95% flat | 未出现 | 消除 |
|
||||
|
||||
## 8. 剩余优化机会
|
||||
|
||||
- **bufio.Reader/Writer 池化**: heap 中仍占主导，代理路径可优化
|
||||
- **连接池调优**: `net.Dialer.DialContext` 在代理路径仍有分配
|
||||
- **系统调用基线**: syscall 仍占 60%+ CPU，io_uring 可进一步挖掘
|
||||
|
||||
## 9. 原始数据文件
|
||||
|
||||
- `benchmarks/v0.4.0/pprof/v2/cpu-final.prof` — 优化后 CPU profile
|
||||
- `benchmarks/v0.4.0/pprof/v2/allocs-final.prof` — 优化后分配 profile
|
||||
- `benchmarks/v0.4.0/cpu-top-final.txt` — 优化后 CPU top 函数
|
||||
- `benchmarks/v0.4.0/allocs-top-final.txt` — 优化后分配 top 函数
|
||||
- `benchmarks/v0.4.0/pprof/cpu.prof` — 原始 CPU profile（保留）
|
||||
- `benchmarks/v0.4.0/pprof/allocs.prof` — 原始分配 profile（保留）
|
||||
- `benchmarks/v0.4.0/cpu-top.txt` — 原始 CPU top 函数
|
||||
- `benchmarks/v0.4.0/allocs-top.txt` — 原始分配 top 函数
|
||||
- `benchmarks/v0.4.0/heap-top.txt` — 堆内存 top 函数
|
||||
- `benchmarks/v0.4.0/summary.txt` — 基准测试汇总
|
||||
|
||||
112
scripts/bench-compare.sh
Executable file
112
scripts/bench-compare.sh
Executable file
@ -0,0 +1,112 @@
|
||||
#!/usr/bin/env bash
|
||||
# bench-compare.sh — 比较两次基准测试结果，检测性能回归。
|
||||
#
|
||||
# 用法:
|
||||
# ./scripts/bench-compare.sh <old-summary> <new-summary>
|
||||
#
|
||||
# 返回码:
|
||||
# 0 — 无显著回归
|
||||
# 1 — 检测到显著回归（默认阈值: latency +10%, allocs +10%）
|
||||
#
|
||||
# 示例:
|
||||
# ./scripts/bench-compare.sh benchmarks/v0.4.0/summary.txt benchmarks/v0.5.0/summary.txt
|
||||
|
||||
set -euo pipefail
|
||||
|
||||
OLD="${1:-}"
|
||||
NEW="${2:-}"
|
||||
|
||||
if [[ -z "$OLD" || -z "$NEW" ]]; then
|
||||
echo "用法: $0 <old-summary> <new-summary>" >&2
|
||||
exit 2
|
||||
fi
|
||||
|
||||
if [[ ! -f "$OLD" ]]; then
|
||||
echo "错误: 找不到旧摘要文件: $OLD" >&2
|
||||
exit 2
|
||||
fi
|
||||
|
||||
if [[ ! -f "$NEW" ]]; then
|
||||
echo "错误: 找不到新摘要文件: $NEW" >&2
|
||||
exit 2
|
||||
fi
|
||||
|
||||
LATENCY_THRESH="${LATENCY_THRESH:-10.0}"
|
||||
ALLOCS_THRESH="${ALLOCS_THRESH:-10.0}"
|
||||
RPS_THRESH="${RPS_THRESH:--10.0}"
|
||||
|
||||
echo "=== 基准比较 ==="
|
||||
printf "%-40s %12s %12s %12s\n" "Benchmark" "Old" "New" "Change%"
|
||||
echo "----------------------------------------------------------------------"
|
||||
|
||||
REGRESSION=0
|
||||
|
||||
# 提取并比较关键微基准（ns/op 和 B/op）
|
||||
# 格式: BenchmarkName-NN 1234 ns/op 567 B/op 890 allocs/op
|
||||
compare_metric() {
|
||||
local bench="$1"
|
||||
local metric="$2"
|
||||
local thresh="$3"
|
||||
local better_is_lower="${4:-1}"
|
||||
|
||||
local old_val new_val
|
||||
old_val=$(grep -E "^${bench}" "$OLD" | grep -oE "[0-9]+(\.[0-9]+)?[[:space:]]*${metric}" | head -1 | awk '{print $1}')
|
||||
new_val=$(grep -E "^${bench}" "$NEW" | grep -oE "[0-9]+(\.[0-9]+)?[[:space:]]*${metric}" | head -1 | awk '{print $1}')
|
||||
|
||||
if [[ -z "$old_val" || -z "$new_val" ]]; then
|
||||
return 0
|
||||
fi
|
||||
|
||||
if awk -v o="$old_val" 'BEGIN { exit (o == 0) ? 0 : 1 }'; then
|
||||
printf "%-40s %12s %12s %11s%%\n" "$bench ($metric)" "$old_val" "$new_val" "N/A"
|
||||
return 0
|
||||
fi
|
||||
|
||||
local change
|
||||
change=$(awk -v o="$old_val" -v n="$new_val" 'BEGIN { printf "%.2f", ((n - o) / o) * 100 }')
|
||||
local abs_change
|
||||
abs_change=$(awk -v c="$change" 'BEGIN { printf "%.2f", c < 0 ? -c : c }')
|
||||
|
||||
printf "%-40s %12s %12s %11s%%\n" "$bench ($metric)" "$old_val" "$new_val" "$change"
|
||||
|
||||
if awk -v c="$abs_change" -v t="$thresh" 'BEGIN { exit (c > t) ? 0 : 1 }'; then
|
||||
if [[ "$better_is_lower" == "1" && $(awk -v c="$change" 'BEGIN { print (c > 0) ? 1 : 0 }') -eq 1 ]]; then
|
||||
echo " ⚠️ 回归警告: $bench $metric 增加 ${change}% (阈值 ${thresh}%)" >&2
|
||||
REGRESSION=1
|
||||
elif [[ "$better_is_lower" == "0" && $(awk -v c="$change" 'BEGIN { print (c < 0) ? 1 : 0 }') -eq 1 ]]; then
|
||||
echo " ⚠️ 回归警告: $bench $metric 降低 ${change}% (阈值 ${thresh}%)" >&2
|
||||
REGRESSION=1
|
||||
fi
|
||||
fi
|
||||
}
|
||||
|
||||
# 关键基准测试前缀列表（前缀匹配）
|
||||
BENCHMARKS=(
|
||||
"BenchmarkAccessLogProcess"
|
||||
"BenchmarkFileCacheGet"
|
||||
"BenchmarkProxyCacheGet"
|
||||
"BenchmarkStaticFile"
|
||||
"BenchmarkStaticIndex"
|
||||
"BenchmarkStaticTryFiles"
|
||||
"BenchmarkProxyForward"
|
||||
"BenchmarkProxyHostClient"
|
||||
"BenchmarkProxyWithMockBackend"
|
||||
"BenchmarkMiddlewareProcessChain"
|
||||
"BenchmarkMiddlewareChainExecution"
|
||||
"BenchmarkCompressionMiddleware"
|
||||
"BenchmarkDNSResolverLookupWithCache"
|
||||
)
|
||||
|
||||
for bench in "${BENCHMARKS[@]}"; do
|
||||
compare_metric "$bench" "ns/op" "$LATENCY_THRESH" 1
|
||||
compare_metric "$bench" "B/op" "$ALLOCS_THRESH" 1
|
||||
done
|
||||
|
||||
echo ""
|
||||
if [[ "$REGRESSION" -eq 0 ]]; then
|
||||
echo "✅ 未检测到显著性能回归"
|
||||
exit 0
|
||||
else
|
||||
echo "❌ 检测到性能回归，请检查上述警告"
|
||||
exit 1
|
||||
fi
|
||||
Loading…
x
Reference in New Issue
Block a user