Linux defaults are deliberately conservative — tuned for laptops and small virtual machines, not a dedicated server with 128 GB of RAM and a 10Gbps uplink. A short, well-understood sysctl configuration can noticeably improve throughput and connection handling. The equally important skill is restraint: half the tuning advice circulating online is outdated cargo cult that modern kernels handle better automatically.

Ground Rules First

  • Change one thing at a time and benchmark before and after — otherwise you are guessing.
  • Persist settings in /etc/sysctl.d/99-tuning.conf, apply with sysctl --system, and keep the file in version control.
  • Modern kernels (6.x, standard on 2026 distributions) auto-tune many buffers; verify a setting is still relevant before copying it from a 2015 blog post.

Network: Where the Real Wins Are

For servers handling many connections or high bandwidth:

net.core.somaxconn = 4096
net.ipv4.tcp_max_syn_backlog = 8192
net.ipv4.ip_local_port_range = 10240 65000
net.core.rmem_max = 67108864
net.core.wmem_max = 67108864
net.ipv4.tcp_congestion_control = bbr
net.core.default_qdisc = fq

The backlog settings prevent connection drops under bursts; the buffer maxima matter for high bandwidth-delay paths (a 10Gbps link to another continent needs large windows); and BBR congestion control is the single most impactful switch for long-distance throughput, often doubling it on lossy paths.

File Descriptors and Connection-Heavy Workloads

A busy proxy or websocket server exhausts the default open-file limits quickly. Raise fs.file-max at the kernel level and per-service limits via systemd (LimitNOFILE=65535). Symptoms of hitting the ceiling — "too many open files" errors — often masquerade as application bugs.

Memory: swappiness and Dirty Ratios

On a database or cache server, set vm.swappiness = 10 so the kernel strongly prefers dropping page cache over swapping application memory. Lowering vm.dirty_ratio and vm.dirty_background_ratio (for example to 10 and 5) smooths write bursts on servers with large RAM, avoiding multi-second stalls when the kernel flushes tens of gigabytes at once. Databases with their own durability guarantees often document specific recommendations — PostgreSQL and Kafka both do; follow the application vendor over generic advice.

What to Leave Alone

Skip tcp_tw_recycle (removed from modern kernels for breaking NAT clients), do not disable swap entirely on general-purpose servers, and avoid pinning huge pages unless your database documentation explicitly calls for it. If a setting has no measurable effect in your benchmark, remove it — every unexplained line in sysctl.conf is future confusion.

Frequently Asked Questions

How much improvement is realistic?

For long-distance transfers, BBR plus larger buffers can double throughput. For local LAN workloads, gains are modest — single-digit percentages. Connection-heavy servers see the biggest wins from backlog and file descriptor limits.

Can bad sysctl values break the server?

Rarely fatally, but yes — absurd buffer sizes can waste RAM and extreme dirty ratios can cause stalls. That is why you change one value at a time and keep the config in Git.

Do these settings apply to VPS too?

Partially. On a VPS the hypervisor limits what you control; on bare metal every knob is yours — one of the quiet advantages of dedicated hardware.

Get hardware worth tuning — explore our high-performance dedicated servers or ask for a performance consultation.