Benchmarks

Latest Benchmark Results

Run date: 2026-03-28T21:50:52Z Git SHA: 1577d67 Node count: 10,000

Latency

The stress test measures three distinct latency phases:

• Pod startup — time from node creation until all expected DaemonSet pods become Ready. This is dominated by simulated pod startup delays and reflects real-world DaemonSet boot time. Excludes never-ready nodes.
• Vigil reaction — time from all pods becoming Ready until the taint is removed. This isolates Vigil’s controller overhead: reconcile loop detection, readiness verification, and the taint removal API call.
• End-to-end — total time from node creation to taint removal. For never-ready nodes, this equals the controller timeout.

Pod Startup Latency

Vigil Reaction Time

End-to-End Latency

Outcomes

Resource Utilization

Resource Utilization Over Time

Heap Memory

xychart-beta
    title "Heap Allocation Over Time"
    x-axis "Elapsed (s)" [5, 40, 75, 110, 145, 180, 215, 250, 285, 320, 355, 390, 425, 460, 495, 530, 565, 600, 635, 670, 705, 740, 775, 810, 845, 880, 915, 950, 985, 1020]
    y-axis "Heap (MB)" 0 --> 280
    line [8, 11, 24, 31, 47, 41, 60, 76, 86, 73, 86, 114, 117, 88, 99, 97, 157, 128, 163, 187, 194, 190, 165, 227, 174, 223, 140, 165, 186, 214]

Goroutines

xychart-beta
    title "Active Goroutines Over Time"
    x-axis "Elapsed (s)" [5, 40, 75, 110, 145, 180, 215, 250, 285, 320, 355, 390, 425, 460, 495, 530, 565, 600, 635, 670, 705, 740, 775, 810, 845, 880, 915, 950, 985, 1020]
    y-axis "Goroutines" 0 --> 250
    line [140, 185, 197, 196, 195, 194, 192, 199, 199, 199, 190, 192, 192, 194, 196, 194, 195, 199, 199, 192, 195, 199, 194, 204, 194, 194, 192, 190, 195, 133]

Memory grows linearly with node count (10 nodes/sec) as the informer cache accumulates Node and Pod objects. The sawtooth pattern in heap allocation shows the Go garbage collector reclaiming memory between allocation bursts. Goroutine count remains stable (~190-200), confirming no goroutine leaks.

Test Configuration

Profile Distribution