Lanbench [top] Link

LANBench: Write-Up

Why Standard Benchmarks Fail (The LANBench Advantage)

Before LANBench gained traction, engineers relied on tools like text-generation-webui’s built-in stats or llama.cpp’s --benchmark flag. These tools have a fatal flaw: they measure internal compute speed.

Consider this scenario:

  • Your AI server processes 100 tokens per second locally.
  • Your network switch is congested.
  • Your client script is written in Python, adding 50ms of serialization overhead.

Standard benchmarks would report "100 t/s." LANBench would report the truth: "65 t/s due to network jitter and JSON parsing." LANBench

How LANBench works (high level)

  1. Test endpoints are deployed across the LAN (physical hosts, VMs, or containers).
  2. A controller schedules test scenarios—defining traffic types, rates, packet sizes, duration, and QoS markings.
  3. Endpoints generate traffic according to scenario parameters; measurements are collected at sender and receiver.
  4. Results are aggregated to compute throughput, round-trip or one-way latency, jitter, packet loss, and retransmission statistics.
  5. Reports visualize time-series behavior and summarize per-test metrics and percentiles (e.g., p50, p95, p99 latency).

4. Burn Testing

It can be used to stress-test network links to check for overheating hardware or unstable connections that drop under heavy load. Your AI server processes 100 tokens per second locally

Product Name: LANBench

Tagline: "The Ultimate Local Area Network Stress Test." Standard benchmarks would report "100 t/s