ykxmz/_build/ab_tcp.sh

#!/bin/bash
# TCP 重传调优的受控 A/B。
#
# 背景（实测）：
#   TCPTimeouts: 3145 · TCPLostRetransmit: 2010
#   这些多秒/多分钟的尾巴全是 TCP 重传超时造成的。
#   这台机器内核是定制的，**没有 BBR**（只有 reno/cubic），
#   也没有任何替代 qdisc（只有 noqueue），所以只能在
#   现有 cubic 上调 TCP 的重传行为。
#
# 三个候选开关（都对症"细流 + 高丢包"这个场景）：
#   tcp_thin_linear_timeouts  细流不搞 RTO 指数退避，重传来得更快
#   tcp_fack                  前向确认，丢包恢复更积极
#   tcp_retries2              重传几次就放弃（8 太大，卡十几分钟）
#
# 每轮：设参数 -> 打 60 次 HTTP -> 记录成功率/耗时分布 -> 下一轮
# 全部跑完回滚，最后按数据决定采用哪组。
set -u

BASE=https://relay.xxmxxs.top
UA='LanRemote/1.8.1'
N=60

# 保存原始值，跑完恢复
declare -A ORIG
for k in tcp_thin_linear_timeouts tcp_fack tcp_retries2 tcp_autocorking; do
    ORIG[$k]=$(sysctl -n net.ipv4.$k 2>/dev/null || echo "")
done

restore() {
    for k in "${!ORIG[@]}"; do
        [ -n "${ORIG[$k]}" ] && sysctl -qw net.ipv4.$k="${ORIG[$k]}" 2>/dev/null || true
    done
}

measure() {
    local label="$1"
    local ok=0 fail=0 times=()
    for i in $(seq 1 "$N"); do
        out=$(curl -s -o /dev/null -w '%{http_code} %{time_total}' --max-time 25 \
              -H "User-Agent: $UA" "$BASE/v1/health")
        c=$(echo "$out" | awk '{print $1}')
        ms=$(echo "$out" | awk '{printf "%d", $2*1000}')
        if [ "$c" = "200" ]; then ok=$((ok+1)); times+=("$ms"); else fail=$((fail+1)); fi
    done
    local pct=$((ok * 100 / N))
    local stats="无数据"
    if [ "${#times[@]}" -gt 0 ]; then
        stats=$(printf '%s\n' "${times[@]}" | sort -n | awk '{a[NR]=$1} END {
            printf "中位 %d  P90 %d  P99 %d  最大 %d", a[int(NR/2)+1], a[int(NR*0.9)+1], a[int(NR*0.99)+1], a[NR]}')
    fi
    printf "  %-34s %3d%% (%d/%d)  %s ms\n" "$label" "$pct" "$ok" "$N" "$stats"
    echo "$label|$pct|$ok|$stats" >> /tmp/tcpab.txt
}

rm -f /tmp/tcpab.txt

echo "════════ 原始值 ════════"
for k in "${!ORIG[@]}"; do printf "  %-30s = %s\n" "$k" "${ORIG[$k]}"; done
echo

echo "════════ 各组合实测（每轮 $N 次）════════"

echo "  [1/6] 基线（不改）"
measure "基线"

echo "  [2/6] thin_linear_timeouts=1"
sysctl -qw net.ipv4.tcp_thin_linear_timeouts=1
measure "thin=1"

echo "  [3/6] thin=1 + fack=1"
sysctl -qw net.ipv4.tcp_fack=1
measure "thin=1 fack=1"

echo "  [4/6] thin=1 fack=1 + retries2=5"
sysctl -qw net.ipv4.tcp_retries2=5
measure "thin=1 fack=1 retries2=5"

echo "  [5/6] thin=1 fack=1 retries2=5 + autocorking=0"
sysctl -qw net.ipv4.tcp_autocorking=0
measure "全部开启"

echo "  [6/6] 回滚到基线再测一次（排除时间漂移）"
restore
measure "回滚后基线"

echo
echo "════════ 汇总（按可用率排序）════════"
sort -t'|' -k2 -rn /tmp/tcpab.txt | while IFS='|' read -r label pct ok stats; do
    printf "  %-34s %3s%%   %s\n" "$label" "$pct" "$stats"
done

echo
echo "════════ 按数据决定 ════════"
BEST=$(sort -t'|' -k2 -rn /tmp/tcpab.txt | head -1)
BEST_LABEL=$(echo "$BEST" | cut -d'|' -f1)
BEST_PCT=$(echo "$BEST" | cut -d'|' -f2)
echo "  最好的是: $BEST_LABEL ($BEST_PCT%)"

case "$BEST_LABEL" in
    "基线"|"回滚后基线")
        echo "  → 调优没有收益，保持默认（只留已验证过的 keepalive 那几项）"
        restore
        ;;
    "thin=1")
        sysctl -qw net.ipv4.tcp_thin_linear_timeouts=1
        echo "  → 只启用 thin_linear_timeouts"
        ;;
    "thin=1 fack=1")
        sysctl -qw net.ipv4.tcp_thin_linear_timeouts=1
        sysctl -qw net.ipv4.tcp_fack=1
        echo "  → 启用 thin_linear_timeouts + fack"
        ;;
    *)
        sysctl -qw net.ipv4.tcp_thin_linear_timeouts=1
        sysctl -qw net.ipv4.tcp_fack=1
        sysctl -qw net.ipv4.tcp_retries2=5
        sysctl -qw net.ipv4.tcp_autocorking=0
        echo "  → 全部启用"
        ;;
esac

echo
echo "  当前生效值:"
for k in tcp_thin_linear_timeouts tcp_fack tcp_retries2 tcp_autocorking; do
    printf "    %-30s = %s\n" "$k" "$(sysctl -n net.ipv4.$k 2>/dev/null)"
done