nsys-analyze

name

nsys-analyze

description

Analyze profiling data from Kit-based apps. Covers Omniverse-specific NVTX zone interpretation, phase detection using sqlite3, Tracy Statistics/Range Limit analysis, csvexport fallback queries, and two-version comparison methodology. Use after capturing profiles with the profiling skill. NOT for capturing traces (use profiling), adding zones to code (use profiling-api), or applying fixes (use perf-tuning).

Profile Analysis for Omniverse / Kit-based Apps

Analyze profiling data from Kit, Isaac Sim, and Isaac Lab using sqlite3 (for .nsys-rep), Tracy Statistics/Range Limit (primary .tracy path), and csvexport (automated .tracy fallback). For capturing profiles and installing tools, see the profiling and install-profilers skills.

Required tools: nsys, sqlite3, csvexport; Tracy GUI is needed for the primary .tracy Statistics workflow. See install-profilers skill.

Omniverse NVTX Zone Reference

Zone Pattern	Meaning	Phase
App Update / App Main loop	Frame boundaries	Runtime
UsdFileOp / UsdFileOp::open / UsdFileOp::newStage	Stage operations	Startup/Loading
UsdContext::Impl::render	USD render context	Runtime
RtxHydraEngine::render*	RTX render passes	Runtime
Hydra render views*	Hydra render delegate ops	Runtime
OmniGraph::* / ComputeGraphImpl::*	OmniGraph compute	Runtime
GeoTreeNode::* / Fabric::*	Fabric/scene population	Loading/Runtime
Carbonite::* / carb::*	Low-level framework (noise — exclude)	All
Thread waiting...	Idle thread (noise — exclude)	All
Executing task / Running fiber	Task scheduler (noise — exclude)	All

Phase Detection Rules

Kit apps have phases: startup → loading → runtime → shutdown.

• Startup = trace start → first App Update frame
• Loading frames = frames with duration > 5× median (stage loading spikes — can appear at start or mid-run)
• Runtime frames = frames with duration ≤ 5× median (steady-state)
• Frame marker = App Update zone (NOT App::beginUpdate)

Note: Loading in Kit apps often happens during runtime as a long frame, not as a separate phase before the first frame. The 5× median threshold reliably separates loading spikes from runtime frames.

---

Analysis Path A: nsys SQLite (for .nsys-rep files)

Step 1: Export to SQLite

nsys export --type=sqlite -o profile.sqlite profile.nsys-rep --force-overwrite=true

Step 2: Overview + Phases + Frame Analysis

sqlite3 -header -column profile.sqlite "
WITH frames AS (
  SELECT ROW_NUMBER() OVER (ORDER BY e.start) as n,
         e.start, e.end, (e.end - e.start) as dur_ns
  FROM NVTX_EVENTS e LEFT JOIN StringIds s ON e.textId = s.id
  WHERE COALESCE(e.text, s.value) = 'App Update' AND e.end IS NOT NULL
),
frame_med AS (
  SELECT dur_ns as med FROM frames ORDER BY dur_ns
  LIMIT 1 OFFSET (SELECT COUNT(*)/2 FROM frames)
),
runtime AS (
  SELECT dur_ns FROM frames, frame_med WHERE dur_ns <= med * 5 ORDER BY dur_ns
)
SELECT
  ROUND((SELECT (MIN(start) - (SELECT MIN(start) FROM NVTX_EVENTS)) / 1e9 FROM frames), 2) as startup_sec,
  ROUND((SELECT (MAX(end) - MIN(start)) / 1e9 FROM frames), 2) as total_sec,
  (SELECT COUNT(*) FROM frames) as total_frames,
  (SELECT COUNT(*) FROM frames, frame_med WHERE dur_ns > med * 5) as loading_frames,
  COUNT(*) as runtime_frames,
  ROUND(AVG(dur_ns)/1e6, 2) as mean_ms,
  (SELECT ROUND(dur_ns/1e6,2) FROM runtime LIMIT 1 OFFSET (SELECT COUNT(*)/2 FROM runtime)) as p50_ms,
  (SELECT ROUND(dur_ns/1e6,2) FROM runtime LIMIT 1 OFFSET (SELECT CAST(COUNT(*)*0.95 AS INT) FROM runtime)) as p95_ms,
  ROUND(MIN(dur_ns)/1e6, 2) as min_ms,
  ROUND(MAX(dur_ns)/1e6, 2) as max_ms,
  ROUND(1000.0/(AVG(dur_ns)/1e6), 1) as fps
FROM runtime;
"

Step 3: Top Zones (runtime only, noise excluded)

sqlite3 -header -column profile.sqlite "
WITH frames AS (
  SELECT ROW_NUMBER() OVER (ORDER BY e.start) as n,
         e.start, e.end, (e.end - e.start) as dur_ns
  FROM NVTX_EVENTS e LEFT JOIN StringIds s ON e.textId = s.id
  WHERE COALESCE(e.text, s.value) = 'App Update' AND e.end IS NOT NULL
),
frame_med AS (
  SELECT dur_ns as med FROM frames ORDER BY dur_ns
  LIMIT 1 OFFSET (SELECT COUNT(*)/2 FROM frames)
),
runtime_frames AS (
  -- Keep only frames classified as steady-state runtime. Do not collapse to
  -- one min/max span, because loading spikes can occur between runtime frames.
  SELECT f.start, f.end
  FROM frames f, frame_med m
  WHERE f.dur_ns <= m.med * 5
)
SELECT
  COALESCE(e.text, s.value) as zone_name,
  COUNT(*) as cnt,
  ROUND(AVG(e.end - e.start)/1e6, 3) as avg_ms,
  ROUND(SUM(e.end - e.start)/1e6, 2) as total_ms,
  ROUND(MAX(e.end - e.start)/1e6, 3) as max_ms
FROM NVTX_EVENTS e
LEFT JOIN StringIds s ON e.textId = s.id
WHERE EXISTS (
    SELECT 1 FROM runtime_frames rf
    WHERE e.start >= rf.start AND e.start < rf.end
  )
  AND e.end IS NOT NULL AND (e.end - e.start) > 0
  AND COALESCE(e.text, s.value) NOT LIKE '%Thread waiting%'
  AND COALESCE(e.text, s.value) NOT LIKE 'Carbonite::%'
  AND COALESCE(e.text, s.value) NOT LIKE 'carb::%'
  AND COALESCE(e.text, s.value) NOT IN ('Executing task','Running fiber')
GROUP BY zone_name
HAVING total_ms > 1
ORDER BY total_ms DESC LIMIT 30;
"

SQLite Schema Quick Reference

Table	Use
NVTX_EVENTS	NVTX ranges/markers. No name column — use text (inline) or join textId→StringIds.id.
StringIds	String lookup (id → value)
CUPTI_ACTIVITY_KIND_KERNEL	CUDA kernel launches (empty for Kit/RTX apps — normal)
TARGET_INFO_GPU	GPU hardware info
TARGET_INFO_SYSTEM_ENV	System environment

---

Analysis Path B: Tracy Statistics (primary path for .tracy files)

Use Tracy GUI Statistics for .tracy files when the goal is hotspot ranking, regression analysis, or optimization comparison.

1. Open the .tracy file in Tracy Profiler.
2. Open View -> Statistics.
3. Drag-select the steady-state interval on the timeline and set Range Limit.
4. Record Mean, Median, Min, Max, Std Dev, Count, and Total Time for key zones.
5. For before/after comparisons, use the same hardware, scene, parameters, and equal-length steady-state Range Limits.

Do not compare a single frame unless the issue is known to occur in one frame and is reproduced across multiple runs.

---

Analysis Path C: Tracy CSV (automated fallback for .tracy files)

csvexport profile.tracy > zones.csv

Inspect the header before scripting against csvexport output. Tracy versions and builds can differ:

• Guide examples use name, mean, count, and total_time.
• Other builds emit nanosecond-specific names such as total_ns, counts, and mean_ns.

Normalize the column names in scripts instead of assuming one schema.

Data is pre-aggregated — one row per unique zone, covering the entire trace (no phase separation).

head -1 zones.csv

Tracy CSV limitation: No per-invocation timestamps — only aggregates. For phase-aware analysis, prefer the nsys SQLite path.

---

Two-Version Comparison

With nsys SQLite (recommended)

nsys export --type=sqlite -o v1.sqlite v1.nsys-rep --force-overwrite=true
nsys export --type=sqlite -o v2.sqlite v2.nsys-rep --force-overwrite=true

Run the overview/frames/zones queries (Steps 2-3) on both databases, save outputs, then compare.

With Tracy CSV

csvexport v1.tracy > v1_zones.csv
csvexport v2.tracy > v2_zones.csv

Compare with Python:

import csv

def number(row, *names):
    for name in names:
        value = row.get(name)
        if value not in (None, ""):
            return float(value)
    return 0.0

def load_zones(path):
    zones = {}
    with open(path) as f:
        for row in csv.DictReader(f):
            name = row.get('name') or row.get('zone_name')
            if not name:
                continue
            zones[name] = {
                'total_ms': number(row, 'total_ns', 'total_time') / 1e6,
                'count': int(number(row, 'counts', 'count')),
                'mean_ms': number(row, 'mean_ns', 'mean') / 1e6,
            }
    return zones

v1, v2 = load_zones('v1_zones.csv'), load_zones('v2_zones.csv')

diffs = []
for name in set(v1) | set(v2):
    t1 = v1.get(name, {}).get('total_ms', 0)
    t2 = v2.get(name, {}).get(, )
     t1 >   t2 > :  
        diffs.append((name, t1, t2, t2 - t1))

()
 name, t1, t2, d  (diffs, key= x: -x[])[:]:
    ()

()
 name, t1, t2, d  (diffs, key= x: x[])[:]:
    ()

Report Structure

1. Overall metrics — total duration, frame count per version
2. Phase comparison — startup time, loading frames count/duration
3. Frame analysis — mean frametime, P50, P95, FPS (runtime frames only)
4. Top regressions — zones slower in v2, ranked by absolute ms impact
5. Top improvements — zones faster in v2
6. New/removed zones — zones appearing only in one version
7. Root cause analysis — explain why the change happened

The goal: not just "FPS dropped 10%" but "FPS dropped 10% because rtUpdatePipeline added 59ms/frame in v2, a new shader pipeline recompilation step not present in v1."

---