Client Area

CentOS Server Performance Tuning: Optimization Techniques for 2026

CentOS Stream (9 or 10 in 2026) inherits Red Hat Enterprise Linux’s robust tuning ecosystem. With kernel 6.12+ in Stream 10, modern hardware support (including x86-64-v3 baseline), and continuous updates, the focus shifts toward balanced throughput, low latency, energy-aware scaling, and workload-specific profiles rather than extreme micro-optimizations.

Performance tuning remains iterative: measure baselines (with tools like perf, sar, tuned-adm profile list, stress-ng, fio, iperf3, or wrk), apply changes, re-measure, and revert if regressions appear. Avoid blanket “max performance” settings — they often hurt latency or stability under mixed loads.

1. Use Tuned Profiles – The Smart Starting Point

tuned is the official dynamic tuning daemon. It applies sysctl, CPU governor, disk elevator, power, and IRQ changes based on workload.

Install & list profiles:

Bash
sudo dnf install tuned
sudo tuned-adm list

Recommended 2026 profiles:

  • throughput-performance — Best default for general servers (web, API, file serving). Disables power saving, uses deadline/noop scheduler, boosts network/disk sysctls.
  • network-throughput — For high-bandwidth NICs (40/100 Gbps+); emphasizes large buffers and interrupt coalescing.
  • latency-performance — For low-jitter needs (databases, real-time apps); pins CPU to performance governor, minimizes C-states.
  • virtual-guest / virtual-host — For KVM/VMware guests or hosts.
  • balanced — Default; good compromise if unsure.

Activate:

Bash
sudo tuned-adm profile throughput-performance
sudo tuned-adm active

Monitor: tuned-adm recommend suggests based on current usage.

2. CPU & Power Management

Modern servers throttle aggressively for power. Disable for consistent performance:

  • Set governor to performance (or use tuned profile above):

    Bash
    cpupower frequency-set -g performance   # install cpupower if needed

  • BIOS/UEFI: Enable Performance mode, disable C-states below C1, disable power-saving features (P-states, turbo on demand if latency-critical).

  • Disable transparent huge pages if they cause latency spikes (common in databases):

    Bash
    echo never > /sys/kernel/mm/transparent_hugepage/enabled
echo never > /sys/kernel/mm/transparent_hugepage/defrag

    Make permanent in /etc/sysctl.d/99-performance.conf.

3. Memory & VM Tuning

Key sysctls for high-concurrency servers:

  • Reduce swappiness (avoid swapping under memory pressure):

    vm.swappiness = 10 (or 1 for latency-sensitive)

  • Increase file handles & inotify watches:

    fs.file-max = 2097152 fs.inotify.max_user_watches = 524288

  • TCP memory buffers (for high-bandwidth connections):

    net.core.rmem_max = 16777216 net.core.wmem_max = 16777216 net.ipv4.tcp_rmem = 4096 87380 16777216 net.ipv4.tcp_wmem = 4096 65536 16777216

  • Overcommit policy (allow more allocations than RAM):

    vm.overcommit_memory = 1 (careful with OOM killer)

Apply in /etc/sysctl.d/99-performance.conf then sysctl –system.

4. Network Stack Optimizations

For 10/25/100 Gbps+ links:

  • Increase socket backlog & queues:

    net.core.somaxconn = 65535 net.core.netdev_max_backlog = 5000

  • Enable TCP timestamps & window scaling:

    net.ipv4.tcp_timestamps = 1 net.ipv4.tcp_window_scaling = 1

  • Tune congestion control (BBR often best in 2026):

    net.ipv4.tcp_congestion_control = bbr

  • Disable slow-start after idle:

    net.ipv4.tcp_slow_start_after_idle = 0

  • For very high packet rates: increase ring buffers via ethtool -G eth0 rx 4096 tx 4096 (NIC-dependent).

5. Disk I/O & Filesystem Tuning

  • Scheduler: Use mq-deadline or none (NVMe default) for SSDs; avoid cfq/bfq on servers.

  • Read-ahead: Increase for sequential workloads:

    blockdev –setra 4096 /dev/nvme0n1

  • XFS/ext4 mount options (in /etc/fstab):

    noatime,nodiratime,discard (TRIM on SSDs) For XFS: inode64,allocsize=1m on large filesystems.

  • LVM stripe cache + RAID write-back if hardware RAID.

6. IRQ & NUMA Awareness

  • Pin IRQs to specific cores (avoid sharing with application threads):

    Use irqbalance (default) or manual with set_irq_affinity.sh scripts from vendor drivers.

  • NUMA: Bind processes to nodes local to NIC/SSD:

    numactl –cpunodebind=0 –membind=0 nginx

7. Monitoring & Validation Tools (2026 Essentials)

  • tuned-adm, perf top, bpftrace, bcc-tools (e.g., funccount, biolatency)
  • sar, iotop, nethogs, htop with extended metrics
  • prometheus-node-exporter + Grafana for long-term trends
  • Benchmark: fio (disk), iperf3/netperf (network), stress-ng –cpu –vm (memory/CPU)

Quick Prioritization Table (Workload-Based)

Workload TypeTop Priorities (in order)Expected Gain
High-concurrency web/APItuned throughput, BBR, large TCP buffers, somaxconn+30–80% throughput
Database (MySQL/PostgreSQL)latency-performance, no THP, vm.swappiness=1, deadline scheduler-20–50% p99 latency
File/Storage servernetwork-throughput, large readahead, XFS allocsize+50% sequential I/O
Container/K8s hostvirtual-host profile, hugepages disabled if not using, IRQ pinningBetter pod density
Low-latency (real-time-ish)latency-performance, C-states disabled, TCP timestampsReduced jitter

Start with a tuned profile, baseline with real workload traffic, apply 2–3 sysctls at a time, and always test rollback. Over-tuning often causes regressions under bursty or mixed loads.

Leave a Reply