Skip to content

README.md

Latest commit

 

History

History
81 lines (47 loc) · 5.83 KB

README.md

File metadata and controls

81 lines (47 loc) · 5.83 KB
Raw

PurrNet promotional benchmark charts

4 charts, with 1920 × 1080 PNG and matching SVG exports, from 2026-09-09.

PurrNet 1.23.0-beta.45 · FishNet 4.7.3 · Mirror 96.0.1 · NGO 2.13.2 · Fusion 2.1.2 Stable 2279

These are selected resource comparisons, not an overall netcode ranking. Each chart includes all five netcodes on a linear scale starting at zero. Asterisks and hatching mark overload. Relative percentages use unrounded values and round to the nearest whole percent.

State replication bandwidth

State replication bandwidth. PurrNet’s downstream bandwidth was 34% lower than FishNet across the state replication workloads. NetworkTransform + SyncVars, 100 connections, 20 Hz.

State replication bandwidth. PurrNet’s downstream bandwidth was 34% lower than FishNet across the state replication workloads. NetworkTransform + SyncVars, 100 connections, 20 Hz.

PNG / SVG

Arithmetic mean of server txBytesPerSec. Tests: MoveY, MoveWander, SyncVars. Sessions: 100 connections, 20 Hz. 10 s windows. No idle subtraction. Binary byte units (1 KiB = 1,024 bytes).

One published run. Overloaded rows describe saturated servers and are not used for comparative claims.

Messaging server CPU

Messaging server CPU. PurrNet’s server CPU usage was 53% lower than Mirror and 22% lower than FishNet in the messaging workloads. RPCs: server broadcasts + client inputs, 100 connections, 20 Hz.

Messaging server CPU. PurrNet’s server CPU usage was 53% lower than Mirror and 22% lower than FishNet in the messaging workloads. RPCs: server broadcasts + client inputs, 100 connections, 20 Hz.

PNG / SVG

Arithmetic mean of server cpuPercent. Tests: SendRPC, ClientInput. Sessions: 100 connections, 20 Hz. 10 s windows. No idle subtraction. Binary byte units (1 KiB = 1,024 bytes).

One published run. Overloaded rows describe saturated servers and are not used for comparative claims.

GC allocation across workloads

GC allocation across workloads. PurrNet’s estimated GC allocation rate was 82% lower than FishNet across the active workloads. NetworkTransform + SyncVars + RPCs + spawn/despawn, 100 connections, 20 Hz.

GC allocation across workloads. PurrNet’s estimated GC allocation rate was 82% lower than FishNet across the active workloads. NetworkTransform + SyncVars + RPCs + spawn/despawn, 100 connections, 20 Hz.

PNG / SVG

Arithmetic mean of server gcAllocBytesPerSec. Tests: MoveY, MoveWander, SyncVars, SendRPC, ClientInput, SpawnChurn. Sessions: 100 connections, 20 Hz. 10 s windows. No idle subtraction. Binary byte units (1 KiB = 1,024 bytes).

One published run. Overloaded rows describe saturated servers and are not used for comparative claims. Allocation is estimated from frame-to-frame managed-heap growth, using a running average when collection shrinks the heap.

Bandwidth per added connection

Bandwidth per added connection. PurrNet’s additional downstream bandwidth per connection was 27% lower than FishNet from 10 to 100 connections. NetworkTransform + SyncVars + RPCs + spawn/despawn, 10 → 100 connections, 20 Hz.

Bandwidth per added connection. PurrNet’s additional downstream bandwidth per connection was 27% lower than FishNet from 10 to 100 connections. NetworkTransform + SyncVars + RPCs + spawn/despawn, 10 → 100 connections, 20 Hz.

PNG / SVG

Mean of per-test (100-connection − 10-connection) server downstream / 90. Tests: MoveY, MoveWander, SyncVars, SendRPC, ClientInput, SpawnChurn. Sessions: 10 → 100 connections, 20 Hz. 10 s windows. No idle subtraction. Binary byte units (1 KiB = 1,024 bytes).

One published run. Overloaded rows describe saturated servers and are not used for comparative claims.

Source and calculation

  • Workflow run
  • Exact raw data used for these images
  • Run metadata and original summary
  • Plotted values, claims and skipped charts
  • Raw data SHA-256: cb2ea6c33998830768bc4dc9a6fc98bc88736e33d4680f2066acc960ed6d39dc
  • Renderer checkout revision: 3cc896f0a4199ed76d0cb636c67c7675b0bc69e7. This identifies the code checkout, not the data snapshot.
  • Lower/higher comparisons use 100 × (PurrNet / named competitor − 1). Zero baselines and overloaded comparators receive neutral text.
  • Bandwidth and CPU categories use arithmetic means. General GC uses the six active workloads. Scaling averages their per-test (100 connections − 10 connections) / 90 bandwidth deltas.
  • Idle and Static are excluded. Missing or incomplete workloads are not treated as zero.
  • GC labels distinguish heap-growth estimates from profiler counters. A mixture of those methods makes the GC chart unavailable.
  • Source, units and full workload names are retained in the bundled data; the images use short workload descriptions.

Regenerate

Install the pinned dependencies with python -m pip install -r .github/scripts/requirements-promo.txt, then run from the repository root:

python .github/scripts/render-promo.py --ci --output docs/promo/latest

Use --chart 03-general-gc to select one chart. To reproduce this particular bundle, supply its source-data.json with --data and source-summary.md with --summary. A render replaces the generated chart bundle in its output directory, including removing skipped or unselected images.

--ci skips unsupported comparisons with reasons in this README and chart-data.json; unexpected errors still fail the job. Without --ci, an unavailable comparison raises an error.