Skip to contents

Interactive benchmark metrics with ggWebGL

Source: vignettes/interactive-benchmarks.Rmd
interactive-benchmarks.Rmd

Purpose

ggWebGL includes manual benchmark helpers for measuring browser rendering behavior across representative scene types. These helpers are diagnostic: rows without browser metadata are not complete browser measurements, and fixed performance numbers should not be reported from incomplete rows.

Status. Compact transport, progressive upload metadata, and browser metric helpers are Experimental diagnostics. They support local measurement and smoke testing, not public fixed-FPS claims.

The benchmark scripts are not run automatically by package checks. They write only to an explicit output_dir or tempdir().

Metric Schema

The scene benchmark helper records a stable set of columns so local browser runs can be compared without changing downstream analysis code.

benchmark_scene_columns()
#>  [1] "scene_id"                   "package_version"           
#>  [3] "commit_sha"                 "dataset_size"              
#>  [5] "primitive_counts"           "transport_mode"            
#>  [7] "serialized_bytes"           "artifact_bytes"            
#>  [9] "startup_latency_ms"         "first_interactive_frame_ms"
#> [11] "selection_latency_ms"       "median_frame_time_ms"      
#> [13] "p95_frame_time_ms"          "median_fps"                
#> [15] "p95_fps"                    "memory_used_mb"            
#> [17] "browser"                    "browser_version"           
#> [19] "gpu_renderer"               "device"                    
#> [21] "os"                         "pixel_width"               
#> [23] "pixel_height"               "interaction"               
#> [25] "artifact_html"              "artifact_csv"              
#> [27] "status"                     "created_at"

The most important groups are:

  • scene identity and package metadata;
  • payload and artifact sizes;
  • startup and browser frame timings;
  • browser, device, GPU, and pixel-size metadata;
  • artifact paths and completion status.

Manual Dense Embedding Run

The dense embedding helper builds a deterministic million-point scene using compact point transport, progressive upload metadata, the density_splat shader, and rectangular brush selection.

source(system.file("examples", "htmlwidget", "million-point-embedding.R", package = "ggWebGL"))
run_manual_million_point_embedding()

Scene-Type Metrics

Use benchmark_scene_types() to produce a CSV with one row per scene. Browser metrics are populated only when a browser capture path is explicitly requested and available.

source(system.file("benchmarks", "benchmark-scene-types.R", package = "ggWebGL"))
metrics <- benchmark_scene_types(
  output_dir = tempdir(),
  scenes = c("embedding", "trajectories", "surface_mesh", "workflow"),
  point_count = 1000000L,
  include_browser = FALSE
)
metrics[, c("scene_id", "serialized_bytes", "artifact_bytes", "status")]

Workflow Comparison

A useful smoke check is whether the same data mapping can be inspected both as a static ggplot2 graphic and as a browser-native WebGL widget. The reduced example below uses deterministic in-memory data.

workflow_comparison_plot(800L)

workflow_comparison_widget(800L, height = 360)

Interpreting Results

Rows with status = "browser_skipped" are useful for checking payload size and widget artifact size only. Browser timing fields should be interpreted only when the row includes browser, GPU, device, and pixel-size metadata from the machine that performed the run.