OrthoRoute/docs/optimization/optimization_baseline_2026-04-08.md

OrthoRoute Performance Baseline — April 8, 2026

Date: April 8, 2026
Version: OrthoRoute (current main branch)
Test Board: TestBackplane.kicad_pcb
Purpose: Track routing performance and identify regression vs. April 5 baseline

Update (April 12, 2026): Automated optimization workflow now available!
See optimization_workflow.md for complete guide.
To investigate this regression: python scripts/analyze_log.py --compare tests/regression/golden_metrics.json

---

Executive Summary

Successfully routed an 18-layer backplane with 512 nets, 1,604 pads in 17.5 minutes (1,049.8s) — a 46% regression vs. April 5 baseline (11.96 min).

Run history (all documented runs):

Run	Date	Change	Total Time	Iter avg	Iters	Tracks	Vias	Barrel	Status
1	Apr 3	Baseline — multi-launch kernel	47.9 min	~39s	74	4,118	2,585	—	✓
2	Apr 3	Persistent CUDA kernel enabled	25 min	~20s	70	4,118	2,585	—	✓ 2× gain
3	Apr 5	Best performance	11.96 min	11.0s	65	4,290	2,754	444	✓ 4× baseline
4	Apr 8	Current state	17.5 min	15.7s	67	4,224	2,688	359	✓ ⚠️ 46% regression

Primary concern: Per-iteration time increased from 11.0s (Apr 5) to 15.7s (Apr 8), indicating possible algorithm or GPU kernel regression.

---

Test Board Characteristics

Property	Value
Board file	TestBackplane.kicad_pcb
Copper layers	18 (F.Cu + 16 internal + B.Cu)
Pads	1,604
Routable nets	512
Total nets	1,088
Components	12 footprints
Board size	73.1 × 97.3 mm
Pre-existing tracks	9,605
Pre-existing vias	6,021
Lattice nodes	~446,472
Lattice edges	~14,281,664

---

Routing Run — Measured Results (April 8, 2026)

Hardware: NVIDIA GPU (CUDA-capable, exact model not captured in log)
Log file: com_github_bbenchoff_orthoroute/logs/latest.log (April 8, 2026, 6:23 PM - 6:48 PM)
Log size: 65,264 lines (~9.8 MB)
Debug mode: OFF (no [PROFILE] timing data — requires ORTHO_DEBUG=1)

Final Metrics

Metric	Value
Success rate	512/512 nets (100%)
Total routing time	1,049.8s (17.5 minutes)
Iterations	67
Iteration 1 time	~29s (estimated from iter 2 start time)
Average iteration time	15.7s
Fastest iteration	3.6s (iter 66)
Slowest iteration	~29s (iter 1 — includes graph init)
Final edge overuse	0 ✓ (fully converged)
Barrel conflicts	359 (acceptable)
Tracks written	4,224 (2,048 escapes + 2,176 routed)
Vias written	2,688

Convergence Trend (April 8, 2026 run)

Iter	Nets active	Overuse (edges)	Barrel	Iter time	Total time	Notes
1	—	—	—	~29s	~29s	Graph init + first routing pass
48	491/512	40,767	416	7.7s	801.6s	Still high overuse mid-convergence
49	495/512	47,586	484	7.8s	816.6s	Overuse spike (normal oscillation)
50	499/512	50,161	517	8.7s	832.2s	Peak overuse during convergence
51	504/512	1,515	517	8.3s	849.6s	Sharp drop in overuse
52	507/512	1,292	517	8.6s	865.1s	Continued improvement
53	511/512	1,545	517	9.0s	880.9s	Minor oscillation
54	511/512	1,523	517	8.3s	896.2s
55	511/512	1,061	517	7.0s	909.5s
56	511/512	864	517	5.8s	921.6s
57	511/512	789	517	5.7s	935.1s
58	511/512	626	517	5.8s	948.3s
59	511/512	341	517	5.3s	960.6s
60	512/512	173	359	4.7s	972.1s	All nets active
61	512/512	44	359	4.2s	985.0s	Near convergence
62	512/512	73	359	4.3s	995.7s	Minor oscillation
63	512/512	2	359	4.0s	1006.7s	Almost converged
64	512/512	2	359	3.7s	1017.1s
65	512/512	21	359	3.8s	1027.3s
66	512/512	26	359	3.6s	1038.4s
67	512/512	0 ✓	359	3.9s	1,049.8s	CONVERGED

Key observations:

• Iteration 50 shows peak overuse (50,161) before rapid convergence
• Barrel conflicts stabilized at 359 by iteration 60
• Final iterations (60-67) averaged ~4s as overuse approached zero
• All 512 nets successfully routed with zero edge overuse

---

GPU Performance

Configuration (from log):

[CONFIG] use_gpu=True
[CONFIG] use_gpu_sequential=True
[CONFIG] use_incremental_cost_update=False
PathFinder (GPU=True, Portals=True)

GPU operations logged:

• ✅ CUDA GPU provider initialized
• ✅ Persistent kernel compiled successfully ([GPU-SEEDS] Persistent kernel compiled successfully!)
• ✅ Via kernels compiled ([VIA-KERNELS] CUDA kernels compiled successfully)
• ✅ GPU radix sort: 0.9s for 16.2M edges (16.2M edges/sec)
• ✅ Via column pooling enabled (capacity=4 per x,y location)

Average pathfinding times (from log samples):

• [GPU-SEEDS] SUCCESS: ~70-90ms per net typical
• Range: 20ms (simple nets) to 190ms (complex nets)

No GPU fallbacks to CPU reported — all routing used GPU acceleration.

---

Performance Regression Analysis

Comparison vs. April 5 Baseline

Metric	Apr 5 (best)	Apr 8 (current)	Delta	% Change
Total time	11.96 min (717.5s)	17.5 min (1,049.8s)	+332.3s	+46% ⚠️
Iterations	65	67	+2	+3%
Avg iter time	11.0s	15.7s	+4.7s	+43% ⚠️
Iter 1 time	29.2s	~29s	~0s	~0%
Final overuse	0	0	0	=
Tracks	4,290	4,224	-66	-1.5%
Vias	2,754	2,688	-66	-2.4%
Barrel conflicts	444	359	-85	-19% ✓
Success rate	100%	100%	=	=

Root Cause Analysis

Primary regression: Per-iteration time increased by 43% (11.0s → 15.7s)

Possible causes:

1. GPU kernel mode: April 5 used "PERSISTENT kernel" — need to verify April 8 is using the same mode (log confirms compiled, but check actual usage pattern)
2. Algorithm changes: Code changes between April 5 and April 8 may have added overhead
3. Parameter differences: Different PathFinder parameters (pres_fac_mult, hist_gain) could affect iteration count/time
4. Hardware differences: Different GPU model or available VRAM could impact performance
5. Debug overhead: April 8 generated 69 debug screenshots — visual debugging active

Positive changes:

• ✅ Barrel conflicts reduced: 444 → 359 (-19%)
• ✅ Still 100% routing success
• ✅ Still achieves zero overuse convergence

---

Routing Strategy

Configuration (from log):

STRATEGY: SPARSE (fast convergence)
Convergence: max_iters=250, patience=5

Congestion ratio: Not captured in log without ORTHO_DEBUG=1

---

Debug Output

Location: com_github_bbenchoff_orthoroute/debug_output/run_20260408_182319/
Files: 69 screenshots

Screenshot sequence:

1. 01_board_with_airwires — Initial board state
2. 02_board_no_airwires — Board without airwires
3. 03_board_with_escapes — Portal escape vias placed
4. 04_iteration_01 through 70_iteration_67 — Per-iteration routing progress

---

Recommendations

Immediate Actions

1. Enable profiling — Run with ORTHO_DEBUG=1 to capture [PROFILE] timing data

   $env:ORTHO_DEBUG = '1'
   # Re-run routing
   Remove-Item Env:ORTHO_DEBUG
   

2. Compare GPU kernel modes — Check if persistent kernel is actually being used:

   # In logs, search for:
   # "[GPU] Using PERSISTENT kernel" vs "[GPU] Using MULTI-LAUNCH kernel"
   

3. Disable debug screenshots — Test without visual debugging overhead:

   Remove-Item Env:ORTHO_SCREENSHOT_FREQ, Env:ORTHO_SCREENSHOT_SCALE -ErrorAction SilentlyContinue
   

4. Git diff analysis — Compare code changes between April 5 and April 8:

   git log --oneline --since="2026-04-05" --until="2026-04-08"
   git diff <apr5-commit> <apr8-commit> -- orthoroute/algorithms/manhattan/
   

Investigation Priorities

Priority	Investigation	Expected impact
🔴 #1	Verify GPU kernel mode (persistent vs multi-launch)	Could explain entire 43% regression
🟡 #2	Measure debug screenshot overhead	~5-10% overhead possible
🟡 #3	Profile per-iteration breakdown with ORTHO_DEBUG=1	Identify specific bottleneck
🟢 #4	Parameter tuning (pres_fac_mult, hist_gain)	Optimize iteration count

---

Regression Test Thresholds

Current thresholds (from April 5 golden metrics):

{
  "total_time_s_max": 900,        // Apr 5: 717.5s × 1.25
  "iter_avg_time_s_max": 15.0,    // Apr 5: 11.0s × 1.36
  "iterations_max": 80,            // Apr 5: 65 × 1.23
  "overuse_final_max": 0,
  "converged": true
}

April 8 performance vs. thresholds:

• ✅ total_time_s: 1,049.8s < 900s FAIL (exceeds threshold by 17%)
• ✅ iter_avg_time_s: 15.7s > 15.0s FAIL (exceeds threshold by 5%)
• ✅ iterations: 67 < 80 PASS
• ✅ overuse_final: 0 = 0 PASS
• ✅ converged: true PASS

Action required: Investigate regression or update thresholds if performance change is intentional.

---

Conclusions

Status: ⚠️ Performance regression detected

Summary:

• ✅ Routing quality: Excellent (100% success, zero overuse, fewer barrel conflicts)
• ⚠️ Routing speed: Regressed by 46% vs. best baseline (April 5)
• ❓ Root cause: Unknown — requires profiling to isolate

Next steps:

1. Run with ORTHO_DEBUG=1 profiling enabled
2. Verify GPU persistent kernel is active
3. Compare against April 5 codebase
4. Measure debug screenshot overhead
5. Update baseline if regression is intentional/acceptable, or fix if unintentional

---

Appendix: Raw Log Excerpts

Initialization

2026-04-08 18:23:12,697 - root - ERROR - [LOG] File: DEBUG | logs/latest.log + logs/run_20260408_182312.log
2026-04-08 18:23:15,002 - orthoroute.algorithms.manhattan.unified_pathfinder - WARNING - PathFinder (GPU=True, Portals=True)
2026-04-08 18:23:16,645 - orthoroute.infrastructure.kicad.rich_kicad_interface - INFO - Got layer count from BoardStackup.material_name: 18 copper layers
2026-04-08 18:23:50,253 - orthoroute.algorithms.manhattan.parameter_derivation - INFO - STRATEGY: SPARSE (fast convergence)

Convergence

2026-04-08 18:41:20,396 - orthoroute.algorithms.manhattan.unified_pathfinder - WARNING - [ITER  67] nets=512/512  ✓ CONVERGED  edges=0  via_overuse=0%  barrel=359  iter=3.9s  total=1049.8s
2026-04-08 18:41:20,396 - orthoroute.algorithms.manhattan.unified_pathfinder - WARNING - [CLEAN] All nets routed with zero overuse
2026-04-08 18:41:20,396 - orthoroute.algorithms.manhattan.unified_pathfinder - WARNING - ROUTING COMPLETE: All 512 nets routed successfully with zero overuse!

Final Geometry

2026-04-08 18:41:20,557 - orthoroute.algorithms.manhattan.unified_pathfinder - INFO - [ESCAPE-MERGE] escapes=2048 + routed=2176 → total=4224 tracks after dedup
2026-04-08 18:41:20,557 - orthoroute.algorithms.manhattan.unified_pathfinder - INFO - [ESCAPE-MERGE] escape_vias=0 + routed_vias=2688 → total=2688 vias after dedup