Skip to content

PERFORMANCE.md

Latest commit

 

History

History
65 lines (52 loc) · 2.64 KB

PERFORMANCE.md

File metadata and controls

65 lines (52 loc) · 2.64 KB
Raw

Performance baselines

The site publishes what siltide costs to run. This page is how those numbers were produced, so they can be produced again on a machine that is not the one they came from. Nothing here is estimated: every row is the output of a command in this repository.

The machine
Hardware Apple M5, 10 cores, 24 GB
System macOS 26.6.2
Go 1.26.6
siltide v0.0.37-main.1-4-g16fd798
Accelerators present one Apple GPU

A Linux host with eight NVIDIA cards will not reproduce these numbers. The startup figure in particular moves with how many vendors are installed, because most of it is probing them.

Startup and memory

make build
scripts/perf/startup.sh 20     # median of 20 runs
scripts/perf/memory.sh 30      # the interface, sampled every 2s
scripts/perf/memory.sh 30 --service   # the headless collector
Measurement Result What it covers
First rendered snapshot, simulated fleet 0.03 s process start, one collection, one render, no hardware
First rendered snapshot, this machine 0.52 s the same, plus probing fifteen vendors
of which vendor detection 0.43 s what siltide --diagnose reports on its last line
Resident memory, interface 19 MB 17 devices in view, history buffers filled
Resident memory, headless collector 25 MB --service, serving the API and /metrics

Both memory scripts sample every two seconds and report the median, because the first reading is taken before the history buffers have filled and flatters the number.

The parts, measured separately

make bench
Benchmark Result What it is
BenchmarkFilter 11 µs one keystroke in the filter bar: parse the query, match every device and process in view
BenchmarkRender 0.38 ms redrawing the Overview at 200 by 60 cells
BenchmarkRenderDashboard 0.51 ms the same for the Dashboard
BenchmarkCollect/64 0.60 ms one collection pass over 64 devices, excluding what the vendor tool itself costs
BenchmarkRecord 7.5 µs writing one history point to disk

What is not measured here

  • No comparison against another tool. Running one properly means the same fleet, the same window and the same definition of "started", and none of that has been done. A number without that is an advertisement.
  • Vendor tool cost. BenchmarkCollect measures siltide's work over already-parsed output. On a real NVIDIA host most of a pass is NVML.
  • Anything about a large fleet. These are one machine and a simulated seventeen-device fleet. Numbers from a real rack are welcome as an issue.