diagnose-perf

Performance Diagnosis Guide

Quick triage to identify the most likely performance bottleneck in an Isaac Sim or Isaac Lab workload. This skill does NOT require profiling tools — it uses only nvidia-smi, standard Linux utilities, and Kit config inspection.

For deeper analysis after triage, use the profiling and nsys-analyze skills.

Phase 1 — System Snapshot (no Isaac process needed)

Run these commands and check for red flags:

GPU Info

nvidia-smi -q | grep -E "Product Name|FB Memory Usage|GPU Current Temp|Performance State|Clocks Throttle|Driver Version|CUDA Version|PCIe Generation"

Key fields to check:

CPU Info

# Governor (performance = best for benchmarks)
cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor | sort -u

# Core count
nproc

# Current frequency
lscpu | grep "MHz"

Red flag: Governor is powersave or schedutil — for benchmarks, performance is recommended:

# Requires root; may be read-only in containers
sudo cpupower frequency-set -g performance

Memory

free -h

Red flag: Swap usage > 0 during Isaac runs means system RAM is insufficient.

Phase 2 — Runtime Capture (Isaac process running)

Start the Isaac workload, then capture GPU metrics while it runs:

GPU Monitoring

# Run for 30 seconds alongside the workload
nvidia-smi dmon -s pucm -d 1 -c 30 > /tmp/gpu_monitor.csv

Columns: pwr (watts), gtemp (°C), sm (SM utilization %), mem (memory utilization %), fb (VRAM MB used)

Process-Level Check

# Find Isaac process
ISAAC_PID=$(pgrep -f "isaac-sim\|kit\|python.*isaacsim" | head -1)

# CPU and memory usage
top -bn1 -p $ISAAC_PID | tail -1

# Thread count (high = possible thread contention)
ls /proc/$ISAAC_PID/task 2>/dev/null | wc -l

Quick Frame Timing (if benchmark outputs JSON)

If the user ran a benchmark skill, check the output JSON for FPS:

cat /tmp/benchmark_output/*.json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for phase in data:
    for m in phase.get('measurements', []):
        metric = m.get('name') or m.get('metric', '')
        value = m.get('data', m.get('value'))
        if isinstance(metric, str) and ('fps' in metric.lower() or 'time' in metric.lower()) and value is not None:
            print(f\"{phase['phase_name']}: {metric} = {value:.2f}\")
"

Phase 3 — Bottleneck Classification

Use the GPU monitoring data to classify the bottleneck:

Reading the nvidia-smi dmon Output

# Average SM and memory utilization from capture
awk 'NR>2 && $1!="#" {sm+=$5; mem+=$6; n++} END {printf "Avg SM: %.0f%%  Avg MEM: %.0f%%\n", sm/n, mem/n}' /tmp/gpu_monitor.csv

Decision Tree

Physics vs Rendering (if GPU compute-bound)

Without profiling, check these heuristics:

• Physics-heavy scene (>100 rigid bodies, soft bodies, fluids): likely PhysX-bound
• Camera/lidar-heavy scene (multiple render products): likely render-bound
• Both: profile to separate — use profiling skill with NVTX markers

Phase 4 — Triage Handoff

Do not apply fixes from this skill. Use the bottleneck classification above to choose the next skill:

1. Need likely fixes now: use perf-tuning with the red flags and bottleneck category.
2. Need exact hotspot attribution: use profiling to capture traces, then nsys-analyze.
3. Need a benchmark comparison: use benchmark-isaacsim or benchmark-isaaclab for WARM results.

Common handoff topics for perf-tuning: headless/viewport work, Fabric, debug visualization, CPU governor, RTX quality, PhysX settings, collision geometry, and waitIdle/async rendering.

Triage Report Template

After running Phases 1-3, summarize findings in this format:

## Performance Triage Report

### System
- GPU: [model] ([VRAM] GB) — Driver [version], CUDA [version]
- CPU: [model] × [cores] — Governor: [governor]
- RAM: [total] ([used] used, [swap] swap)
- PCIe: Gen[N]

### Red Flags
- [ ] GPU throttling: [yes/no, reason]
- [ ] CPU governor: [performance/powersave/schedutil]
- [ ] VRAM pressure: [usage %]
- [ ] Swap in use: [yes/no]

### Runtime Metrics (30s average)
- GPU SM utilization: [N]%
- GPU memory utilization: [N]%
- VRAM used: [N] MB / [total] MB
- CPU usage: [N]%

### Bottleneck Classification
**[GPU compute / VRAM bandwidth / VRAM capacity / CPU / Loading / Idle]**

### Recommended Actions
1. [Most impactful fix]
2. [Second fix]
3. [If deeper analysis needed: "Profile with nsys — see profiling skill"]

Kit Settings Reference

For the full performance settings reference (physics, rendering, app loop), see the perf-tuning skill.

Settings can be applied via:

• .kit files in apps/ directory
• CLI: --/setting/path=value
• Python: carb.settings.get_settings().set("/setting/path", value)

When to Escalate to Full Profiling

This triage identifies the category of bottleneck. For specific hotspots, use:

• profiling skill → capture nsys/Tracy traces
• nsys-analyze skill → analyze traces to find exact functions/zones causing slowdowns

Escalate when:

• Triage shows GPU compute-bound but you need to know if it's physics or rendering
• FPS is inexplicably low despite healthy system metrics
• The user needs frame-by-frame timing breakdown
• Comparing performance between two versions or configurations