Performance & Profiling

Hard CPU Budgets, Zero Managed GC Allocations & Profiler Markers

Performance in PERCH is treated as an engineering contract rather than an afterthought. The system is architected to guarantee stable frame pacing and zero garbage collection spikes across high-density agent workloads.

Performance Targets & Results

Workload Metric	Target Limit	Measured Result	Status
Steady-State GC Allocations	0 B / tick	0 B / tick	MET
300 Agents / 600 Slots CPU (p95)	$\le 6.0\text{ ms}$	2.67 ms	MET
Lease Contention Batch Throughput	$\le 1.0\text{ ms}$	0.18 ms	MET
Physics Query Budgeting	Bounded Cap	Enforced via `MaxPhysicsQueriesPerTick`	MET

Profiler Markers & Deep Telemetry

PERCH instruments its internal systems with strongly typed ProfilerMarker scopes, allowing you to isolate exact timings in the Unity Profiler:

[Main Thread]
└── PerchWorld.Tick
    ├── SpatialHash.UpdateMovingPerches
    ├── LeaseManager.HeartbeatTick
    ├── CandidateSelector.ScoreCandidates
    │   ├── SpatialHash.QueryRadiusNonAlloc
    │   └── CandidateSelector.CoarseDistanceCull
    ├── PerchPlanner.EvaluateTrajectories
    │   ├── PerchTrajectory.ComputeHermiteLUT
    │   └── Physics.SphereCastNonAlloc
    └── PerchAgent.DrainEventQueue

Recommended Benchmark Methodology

When measuring performance in your own project:

Always build and run a Standalone Player (Release configuration) rather than profiling inside the Unity Editor with deep profiling overhead.
Measure using actual production meshes, collision layers, and target hardware.
Record test environment metadata: Unity version, platform OS, CPU model, agent count, slot count, fixed timestep, warm-up period, and 95th-percentile frame time.