Linux Kernel Tuning for High-Throughput

Summary

A database server started dropping connections during peak hours (~5000 concurrent connections), suspected to be default kernel limits (somaxconn, file descriptors) rather than the application, so sysctl and kernel parameter tuning were carried out.

System Context

• •Server: Ubuntu 22.04, 16 core / 64GB RAM, running PostgreSQL + connection pooler
• •Symptom: connection refused above ~4500 concurrent connections during peak hours
• •Benchmarked before/after with pgbench using the same load scenario

Checklist

• ☑Ran a pgbench baseline before changing any sysctl values
• ☑Applied one change at a time and re-measured to isolate what actually helped
• ☑Confirmed no more connection refused errors at 5000 concurrent load
• ☑Persisted every changed sysctl value to config (survives reboot)

Commands

sysctl -w net.core.somaxconn=65535

sysctl -w net.ipv4.tcp_max_syn_backlog=65535

sysctl -w vm.swappiness=10

sysctl -w fs.file-max=2097152

echo 'net.core.default_qdisc=fq' >> /etc/sysctl.conf

Conclusion

Kernel tuning provides measurable improvements for network-intensive workloads. Always benchmark before and after changes. Document all modifications — kernel updates can reset sysctl values.