performance-tuning – skill DevOps & Hạ tầng cho agent AI

Performance Tuning

Optimize Linux system performance through kernel parameter tuning, I/O scheduler selection, memory management, CPU governor configuration, and benchmarking. Covers methodology, real sysctl settings, and tool-based validation.

When to Use

• Server experiencing high latency, throughput bottlenecks, or resource exhaustion
• Preparing infrastructure for high-traffic events or load tests
• Tuning a database server, web server, or application host for production
• Diagnosing whether a bottleneck is CPU, memory, disk I/O, or network
• Establishing baseline performance metrics before and after changes
• Configuring kernel parameters for containers, VMs, or bare-metal hosts

Prerequisites

• Root or sudo access on the target system
• sysstat package installed (provides sar, iostat, mpstat)
• linux-tools or perf package for CPU profiling
• Benchmarking tools: fio (disk), sysbench (CPU/memory), iperf3 (network)
• Baseline metrics collected before making any changes

Performance Analysis Methodology

Always follow this order:

1. Collect baseline -- measure current performance with tools

2. Identify bottleneck -- determine if CPU, memory, I/O, or network

3. Change one parameter -- apply a single tuning change

4. Measure impact -- re-run the same benchmark

5. Document -- record the change and its effect

6. Iterate or revert -- keep the change if beneficial, revert if not

System Monitoring Tools

# 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

All tuning values and profiles below are hypotheses, not safe universal presets. Verify each parameter against the installed kernel/driver, record original values and a tested rollback, and apply one measured change at a time. Do not run sysctl --system or persistent service/rule edits as a baseline check.

Sysctl Kernel Parameter Tuning

Network Tuning

# /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

Memory Tuning

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

# Tune swappiness within the installed kernel's documented range; modern kernels may support 0-200
# 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

Applying Sysctl Changes

# 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 Scheduler Configuration

# 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 --setra 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 Governor Configuration

# 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 daemon-reload
systemctl enable --now cpu-governor

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

Benchmarking Tools

fio -- Disk I/O Benchmarking

Writes consume capacity and can disrupt workloads. Use a fresh dedicated scratch directory on an approved test filesystem with sufficient free space and an explicit bounded workload. Never point write tests at raw devices, database files or existing user data. Read tests may create test files unless configured otherwise; inspect fio's actual job plan before execution.

# Install fio
apt install -y fio    # Debian/Ubuntu
dnf install -y fi

… (nội dung đầy đủ có trong gói tải về)