Skip to content

Resources

Saphira benchmarks

Practical, reproducible measurements of real Saphira services and workloads. Every result belongs with the machine, route, workload and configuration that produced it.

Saphira Linux dragon mascot

Why these benchmarks exist

These pages are not artificial “hello world” marketing numbers. Their purpose is to document what a real Saphira installation can do, how it was tested, where the bottlenecks appeared, and how a configuration change moved the result. A number without its conditions is just a headline; a number with its machine, route and workload attached is something someone else can reproduce or challenge.

That discipline shows up in the iperf3 and webDragon results: the former publishes the TCP transcript behind a striking aggregate result, while the latter shows how the routed network path and direct Keeper-to-VM path differ for the same workload. The other Dragons below will be filled in the same way, as real-world figures become available.

Published benchmark work

Each entry records a real Saphira service or infrastructure stack. Pick one to inspect its method and results.

iperf3

Populated

Measured aggregate TCP throughput on Egg, including the 713 Gbit/s IPv6 loopback result and the console output behind it.

webDragon

Populated

nginx HTTPS delivery and a dynamic PHP 8.5 workload served through PHP-FPM on Saphira-D. This is the first fully populated benchmark.

mailDragon

Planned

Managed mail: SMTP submission and delivery, IMAP access, spam filtering and deliverability. Figures land as we measure production mail.

DNSDragon

Planned

Authoritative DNS with DNSSEC signing, zone serving and GSLB-style traffic direction. Query and signing performance will be recorded here.

vpnDragon

Planned

Encrypted tunnels and VPN access, typically on WireGuard. Encrypted throughput and handshake behaviour will be measured here.

databaseDragon

Planned

SQLite, MariaDB and PostgreSQL as managed database hosts. Query, transaction and concurrency figures will be recorded here.

aiDragon

Populated

Agent tooling, MCP surfaces and memory, plus observed native BitNet CPU-inference figures on Hatchling. CUDA-in-nspawn remains untested.

How to read a benchmark

Where a Dragon is marked Planned, the page explains what will be measured and will be populated as we obtain real-world figures. Where one is marked Exception (currently only aiDragon) the page explains why a local benchmark does not apply. We would rather say so plainly than publish a contrived number.