TCP Tuning Modern

Buffers, congestion.

Overview

Modern TCP tuning matches kernel parameters to workload. Default sysctls assume a desktop circa 2010; production server workloads need different settings to hit the network capacity the hardware actually offers.

• Buffers plus congestion. tcp_rmem, tcp_wmem, congestion algorithm; the three knobs that shape throughput.
• BBR congestion control. Modern BBR replaces CUBIC for many workloads; better throughput on lossy or high-RTT paths.
• SO_REUSEPORT. Multi-process listening on the same port; supports modern multi-process servers.
• Connection tracking plus port range. nf_conntrack limits catch NAT-table exhaustion; ip_local_port_range supports high-fan-out outbound.

The approach

The practical approach: BBR where supported, generous buffers, SO_REUSEPORT for multi-process servers, monitor conntrack utilisation. The team’s discipline produces tuned hosts that survive load spikes.

• BBR congestion control. sysctl net.ipv4.tcp_congestion_control=bbr; modern Linux default for high-throughput workloads.
• Generous buffers. tcp_rmem and tcp_wmem max sizes raised; supports high-bandwidth, high-RTT connections.
• SO_REUSEPORT. Application-level support; multi-process listeners share the same port; CPU scaling without proxy.
• Monitor conntrack. Watch nf_conntrack utilisation; saturation produces silent connection drops.
• Document the changes. Per-host tuning committed to the repo; supports investigation and rebuild.

Why this compounds

TCP tuning discipline compounds across hosts. Each tuned parameter produces ongoing performance; the team’s network expertise grows; new hosts inherit the muscle.

• Better network performance. Right tuning supports throughput; the network stack stops being the bottleneck.
• Better resilience. Generous buffers absorb bursts; the host survives traffic spikes the default version would not.
• Better tail latency. BBR reduces queuing; p99 latency improves; SLOs become easier to hit.
• Institutional knowledge. Each tuning teaches the kernel; the team’s network engineering muscle grows.

TCP tuning is an operational discipline that pays off across years. Nova AI Ops integrates with network telemetry, surfaces patterns, and supports the team’s performance discipline.