performance-tuning

性能调优

通过调整内核参数、选择 I/O 调度器、管理内存、配置 CPU 调控器和进行基准测试，优化 Linux 系统性能。涵盖调优方法、实际的 sysctl 设置，以及基于工具的验证。

适用场景

• 服务器出现高延迟、吞吐量瓶颈或资源耗尽
• 为高流量活动或负载测试准备基础设施
• 为生产环境调优数据库服务器、Web 服务器或应用程序主机
• 排查瓶颈究竟来自 CPU、内存、磁盘 I/O 还是网络
• 在更改前后建立性能基线指标
• 为容器、虚拟机或裸机主机配置内核参数

前提条件

• 拥有目标系统的 root 或 sudo 访问权限
• 已安装 sysstat 软件包（提供 sar、iostat、mpstat）
• 已安装用于 CPU 分析的 linux-tools 或 perf 软件包
• 基准测试工具：fio（磁盘）、sysbench（CPU／内存）、iperf3（网络）
• 在进行任何更改前收集基线指标

性能分析方法

始终按以下顺序进行：

1. 收集基线 -- 使用工具测量当前性能
2. 识别瓶颈 -- 确定瓶颈来自 CPU、内存、I/O 还是网络
3. 更改一个参数 -- 应用单项调优更改
4. 测量影响 -- 重新运行相同的基准测试
5. 记录 -- 记录更改及其效果
6. 迭代或回退 -- 如果有益则保留更改，否则回退

系统监控工具

# CPU and process monitoring
top                              # Interactive process viewer
htop                             # Enhanced interactive viewer
mpstat -P ALL 2                  # Per-CPU utilization every 2 seconds
pidstat -u 2                     # Per-process CPU usage

# Memory monitoring
free -h                          # Memory summary
vmstat 2                         # Virtual memory stats every 2 seconds
# Columns: r=runnable, b=blocked, si/so=swap in/out, bi/bo=block I/O

# Disk I/O monitoring
iostat -xz 2                     # Extended disk stats every 2 seconds
# Key columns: %util, await (latency), r/s, w/s
iotop -oP                        # Show processes doing I/O

# Network monitoring
sar -n DEV 2                     # Network interface stats
ss -s                            # Socket summary
nstat                            # Network counters

# CPU profiling (requires perf)
perf top                         # Real-time function-level CPU profiling
perf stat -a sleep 10            # System-wide counters for 10 seconds
perf record -g -a sleep 30       # Record 30 seconds of call stacks
perf report                      # Analyze recorded data

# One-liner: check all major resources
echo "=== CPU ===" && mpstat 1 1 && echo "=== MEM ===" && free -h && echo "=== DISK ===" && iostat -x 1 1 && echo "=== NET ===" && ss -s

Sysctl 内核参数调优

网络调优

# /etc/sysctl.d/60-network-performance.conf

# Increase the maximum socket receive/send buffer sizes
net.core.rmem_max = 134217728
net.core.wmem_max = 134217728
net.core.rmem_default = 1048576
net.core.wmem_default = 1048576

# TCP buffer auto-tuning (min, default, max in bytes)
net.ipv4.tcp_rmem = 4096 1048576 134217728
net.ipv4.tcp_wmem = 4096 1048576 134217728

# Increase connection backlog for high-traffic servers
net.core.somaxconn = 65535
net.ipv4.tcp_max_syn_backlog = 65535
net.core.netdev_max_backlog = 65535

# Enable TCP fast open (client and server)
net.ipv4.tcp_fastopen = 3

# Reuse TIME_WAIT sockets for new connections
net.ipv4.tcp_tw_reuse = 1

# Increase the range of ephemeral ports
net.ipv4.ip_local_port_range = 1024 65535

# TCP keepalive tuning (detect dead connections faster)
net.ipv4.tcp_keepalive_time = 120
net.ipv4.tcp_keepalive_intvl = 30
net.ipv4.tcp_keepalive_probes = 3

# Disable slow start after idle (keeps congestion window open)
net.ipv4.tcp_slow_start_after_idle = 0

# Enable BBR congestion control (requires kernel 4.9+)
net.core.default_qdisc = fq
net.ipv4.tcp_congestion_control = bbr

内存调优

# /etc/sysctl.d/60-memory-performance.conf

# Reduce swappiness (0-100, lower = less swap usage)
# 10 for general servers, 1 for database servers
vm.swappiness = 10

# Dirty page ratios (controls when dirty data is flushed to disk)
# Lower values = more frequent, smaller writes (better for SSDs)
vm.dirty_ratio = 20
vm.dirty_background_ratio = 5

# For large-memory systems writing to fast storage
# vm.dirty_ratio = 40
# vm.dirty_background_ratio = 10

# Increase inotify limits (for apps watching many files)
fs.inotify.max_user_watches = 524288
fs.inotify.max_user_instances = 1024

# Maximum number of open file descriptors system-wide
fs.file-max = 2097152

# Virtual memory overcommit
# 0 = heuristic (default), 1 = always overcommit, 2 = never overcommit
vm.overcommit_memory = 0

# For Redis or similar in-memory stores, use:
# vm.overcommit_memory = 1

# Disable Transparent Huge Pages if it causes latency spikes (common with databases)
# Done via boot parameter or runtime:
# echo madvise > /sys/kernel/mm/transparent_hugepage/enabled

应用 Sysctl 更改

# Apply all sysctl files
sysctl --system

# Apply a specific file
sysctl -p /etc/sysctl.d/60-network-performance.conf

# Set a parameter temporarily (lost on reboot)
sysctl -w vm.swappiness=10

# Verify a parameter
sysctl vm.swappiness
sysctl net.ipv4.tcp_congestion_control

I/O 调度器配置

# Check the current scheduler for a device
cat /sys/block/sda/queue/scheduler
# Output example: [mq-deadline] none kyber bfq

# Set the scheduler temporarily
echo mq-deadline > /sys/block/sda/queue/scheduler   # Good for databases
echo none > /sys/block/nvme0n1/queue/scheduler       # Best for NVMe SSDs
echo bfq > /sys/block/sda/queue/scheduler            # Good for interactive desktop

# Scheduler recommendations:
# NVMe SSD:  none (noop)    -- minimal overhead, hardware handles scheduling
# SATA SSD:  mq-deadline    -- provides fairness with low latency
# HDD:       mq-deadline    -- prevents starvation, good for databases
# Desktop:   bfq            -- prioritizes interactive I/O

# Make persistent via udev rule
cat <<'EOF' > /etc/udev/rules.d/60-io-scheduler.rules
# Set mq-deadline for rotational (HDD) devices
ACTION=="add|change", KERNEL=="sd[a-z]", ATTR{queue/rotational}=="1", ATTR{queue/scheduler}="mq-deadline"
# Set none for non-rotational (SSD/NVMe) devices
ACTION=="add|change", KERNEL=="sd[a-z]", ATTR{queue/rotational}=="0", ATTR{queue/scheduler}="none"
ACTION=="add|change", KERNEL=="nvme[0-9]*", ATTR{queue/scheduler}="none"
EOF

udevadm control --reload-rules

# Tune read-ahead for sequential workloads (database sequential scans)
blockdev --setrahead 4096 /dev/sda    # 4096 sectors = 2 MB

# Enable TRIM for SSDs (weekly via systemd timer)
systemctl enable --now fstrim.timer
fstrim -av    # Manual run

CPU 调控器配置

# Check available governors
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_governors
# Output: performance powersave schedutil

# Check current governor
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

# Set all CPUs to performance mode (maximum frequency)
for cpu in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
  echo performance > "$cpu"
done

# Set using cpupower (if installed)
cpupower frequency-set -g performance

# Governor recommendations:
# Server (production):   performance    -- max frequency, lowest latency
# Server (general):      schedutil      -- kernel-driven dynamic scaling
# Laptop / idle server:  powersave      -- minimize power consumption

# Make persistent via systemd service
cat <<'EOF' > /etc/systemd/system/cpu-governor.service
[Unit]
Description=Set CPU governor to performance

[Service]
Type=oneshot
ExecStart=/usr/bin/cpupower frequency-set -g performance
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target
EOF

systemctl enable --now cpu-governor

# Disable CPU boost (turbo) if consistent latency is needed
echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo

基准测试工具

fio -- 磁盘 I/O 基准测试

# Install fio
apt install -y fio    # Debian/Ubuntu
dnf install -y fio    # RHEL/CentOS

# Sequential read test (simulates backup reads)
fio --name=seq-read --ioengine=libaio --direct=1 --rw=read \
  --bs=1M --numjobs=4 --size=1G --runtime=60 --time_based --group_reporting

# Sequential write test
fio --name=seq-write --ioengine=libaio --direct=1 --rw=write \
  --bs=1M --numjobs=4 --size=1G --runtime=60 --time_based --group_reporting

# Random read (4K blocks -- simulates database IOPS)
fio --name=rand-read --ioengine=libaio --direct=1 --rw=randread \
  --bs=4k --numjobs=16 --iodepth=64 --size=1G --runtime=60 --time_based --group_reporting

# Random write (4K blocks)
fio --name=rand-write --ioengine=libaio --direct=1 --rw=randwrite \
  --bs=4k --numjobs=16 --iodepth=64 --size=1G --runtime=60 --time_based --group_reporting

# Mixed random read/write (70/30 -- typical database workload)
fio --name=mixed --ioengine=libaio --direct=1 --rw=randrw --rwmixread=70 \
  --bs=4k --numjobs=8 --iodepth=32 --size=1G --runtime=60 --time_based --group_reporting

sysbench -- CPU 和内存基准测试

# Install: apt install -y sysbench (Debian) / dnf install -y sysbench (RHEL)

# CPU benchmark
sysbench cpu --threads=4 --time=30 run

# Memory benchmark
sysbench memory --threads=4 --time=30 --memory-block-size=1K --memory-total-size=100G run

iperf3 -- 网络基准测试

# Install iperf3
apt install -y iperf3

# Start server on one host
iperf3 -s

# Run client test from another host
iperf3 -c <server-ip> -t 30 -P 4    # 30 seconds, 4 parallel streams

# Test with UDP (measure packet loss)
iperf3 -c <server-ip> -u -b 1G -t 30

# Reverse mode (server sends to client)
iperf3 -c <server-ip> -R -t 30

调优配置快速参考

Web 服务器（nginx/Apache）

# /etc/sysctl.d/60-webserver.conf
net.core.somaxconn = 65535
net.ipv4.tcp_max_syn_backlog = 65535
net.ipv4.tcp_tw_reuse = 1
net.ipv4.tcp_fastopen = 3
net.ipv4.ip_local_port_range = 1024 65535
net.core.default_qdisc = fq
net.ipv4.tcp_congestion_control = bbr
fs.file-max = 2097152
vm.swappiness = 10

数据库服务器（PostgreSQL/MySQL）

# /etc/sysctl.d/60-database.conf
vm.swappiness = 1
vm.dirty_ratio = 15
vm.dirty_background_ratio = 3
vm.overcommit_memory = 2
vm.overcommit_ratio = 80
net.core.somaxconn = 4096
fs.file-max = 2097152
# Disable THP for databases
# echo never > /sys/kernel/mm/transparent_hugepage/enabled

故障排查

症状	诊断命令	常见修复方法
CPU 使用率高，但没有明显的进程	perf top、mpstat -P ALL 2	检查内核级问题：软中断、硬件中断
负载平均值高，CPU 使用率低	vmstat 2（检查 b 列）	I/O 瓶颈：调优调度器，检查磁盘健康状况
系统频繁进行交换	free -h、vmstat 2（检查 si/so）	降低 vm.swappiness，增加 RAM，查找内存泄漏
磁盘延迟突增	iostat -x 2（检查 await）	切换 I/O 调度器，降低脏数据比例，添加 SSD
“打开的文件过多”错误	cat /proc/sys/fs/file-nr	增大 fs.file-max 和 LimitNOFILE
网络吞吐量低	iperf3 -c <server>、ethtool	增大缓冲区，启用 BBR，检查 MTU
负载下应用程序超时	ss -s、sysctl net.core.somaxconn	增大 somaxconn 和 tcp_max_syn_backlog
延迟不稳定	检查 CPU 调控器	将调控器设为 performance，禁用睿频

相关技能

• linux-administration -- 通用系统监控与管理
• systemd-services -- 通过单元文件中的 cgroups 设置资源限制
• block-storage -- 存储层性能（LVM、RAID、文件系统）
• nfs-storage -- NFS 专项性能调优