kit-app-streaming-debug

作成者: nvidia

Kitアプリのライブストリーミングパフォーマンスのボトルネック、WebRTC/ネイティブStreamSDKの遅延、フリーズ、フレームドロップ、ブラウザのWebRTC統計、copy fenceなどを調査する際に使用します。

npx skills add https://github.com/nvidia/omniperf --skill kit-app-streaming-debug

Kit App Streaming Debug

Debug Kit livestream performance end-to-end before jumping into generic profiling. Always collect both server-side Kit logs and browser-side WebRTC evidence, then correlate by timestamp around the lag/freeze event.

Collect Kit Evidence

Run the Kit app with StreamSDK event tracing and verbose file logs. Keep --no-window for app-stream tests when resolution mismatch is suspected.

# Windows
.\repo.bat launch -- --no-window --/exts/omni.kit.livestream.app/primaryStream/enableEventTracing=true --/log/channels/omni.kit.livestream.streamsdk=verbose --/log/fileLogLevel=verbose

# Linux
./repo.sh launch -- --no-window --/exts/omni.kit.livestream.app/primaryStream/enableEventTracing=true --/log/channels/omni.kit.livestream.streamsdk=verbose --/log/fileLogLevel=verbose

If extension context is needed, also enable:

--/log/channels/omni.kit.livestream.webrtc=verbose
--/log/channels/omni.kit.livestream.app=verbose

Extract the known streaming failure markers:

rg -n -C 3 "Timeout of [0-9]+ms exceeded waiting for copy fence|Cannot stream video frame because the video stream is not connected|Still cannot stream video frame|Cannot stream video frame with resolution|NVST_R_BUSY|Client disconnected from WebRTC server" path/to/kit.log

Interpret them this way:

MarkerLikely causeNext check
Timeout of 1000ms exceeded waiting for copy fenceRenderer/GPU stall before the framebuffer copy completesCorrelate with frame time, GPU utilization, and Tracy/NSight render zones
Cannot stream video frame because the video stream is not connectedWebRTC session has not established, or video stream droppedCheck browser ICE state, connection events, and nearby disconnect logs
Still cannot stream video frame...Connection did not recover inside the wait timeout, frame droppedTreat as stream/session instability, not render cost by itself
Cannot stream video frame with resolution ... differs...OS/window resize or AOV/render resolution changed after connectTry --no-window, fixed resolution, disable viewport fill behavior, or use dynamic resize only if required
NVST_R_BUSYStreamSDK is not accepting frames fast enoughIf not adjacent to disconnect/reconnect, suspect encoder, network, or client backpressure
Client disconnected from WebRTC serverClient/session droppedAlign with browser ICE/NACK/PLI/freeze events

Collect Browser Evidence

Open chrome://webrtc-internals before connecting to the stream. After a lag event, use Create dump and inspect packet loss, jitter, frames dropped, decode time, bitrate, freezes, NACK/PLI, and the selected ICE candidate pair.

For ~/proj/ov-web-rtc, run the dev client with npm run dev and capture console logs. Add or inspect an onStreamStats callback like:

onStreamStats: (message: StatsEvent) => {
  const s = message.data.stats;
  console.table({
    fps: s.fps,
    rtdMs: s.rtd,
    decodeMs: s.avgDecodeTime,
    frameLoss: s.frameLoss,
    packetLoss: s.packetLoss,
    bandwidthMbps: s.totalBandwidth,
    bitrateMbps: s.currentBitrate,
    utilizedPct: s.utilizedBandwidth,
    resolution: `${s.streamingResolutionWidth}x${s.streamingResolutionHeight}`,
    codec: s.codec,
  });
}

Browser-side interpretation:

SymptomLikely bottleneck
High avgDecodeTime, low packet lossClient decode/GPU/display bottleneck
High packetLoss, frameLoss, NACK/PLI, jitter, or freezesNetwork or congestion problem
utilizedBandwidth near 100%, bitrate near/above available bandwidthBandwidth cap or congestion control
Low browser FPS with clean browser stats but Kit copy-fence/NVST warningsServer render/copy/encoder/backpressure
Browser resolution differs from expected Kit stream resolutionResize/configuration mismatch

Profiling Follow-Up

Use Tracy only after logs indicate the server side is involved. Add the profiling args to the same Kit run:

--/app/profilerBackend=tracy --/app/profileFromStart=true --/profiler/gpu/tracyInject/enabled=true --/app/profilerMask=1 --/plugins/carb.profiler-tracy.plugin/fibersAsThreads=false --/profiler/channels/carb.events/enabled=false --/profiler/channels/carb.tasking/enabled=false --/rtx/addTileGpuAnnotations=true --/profiler/enabled=true --/profiler/gpu=true --enable omni.kit.profiler.window

Use the profiling skill for capture mechanics and nsys-analyze for trace analysis. Do not promise StreamSDK-internal zones when StreamSDK source is unavailable; correlate log timestamps with surrounding Kit zones such as SharedFrameBuffer::waitForCopy, SharedFrameBuffer::streamBuffer, render zones, and frame markers.

Report Shape

Summarize:

  1. Kit run args and log path.
  2. Browser artifacts: console stats and webrtc-internals dump.
  3. Timeline around the lag event.
  4. Matched markers and likely bottleneck category: renderer/GPU, session/network, resize/config, StreamSDK/encoder backpressure, or client decode.
  5. Next action: config fix, network/client investigation, or server-side profiling.

Source references in this workspace: ~/proj/kit-livestream for omni.kit.livestream.app and omni.kit.livestream.webrtc; ~/proj/ov-web-rtc for browser stats and dev client logging.

nvidiaのその他のスキル

compileiq-debug
nvidia
何かがおかしいときに使用:Search()がハングする、すべての評価がINVALID_SCOREを返す、スコアが改善しない、すべての設定が同じ数値を返す、ptxasエラー…
create-github-pr
nvidia
gh CLIを使用してGitHubのプルリクエストを作成します。ユーザーが新しいPRを作成したい、コードをレビューに提出したい、またはプルリクエストを開きたい場合に使用します。トリガーキーワード -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
他のオープンなIssueをスキャンし、特定のPRが修正する可能性があるものや、誤って壊す可能性があるものを見つけます。隣接修正の機会や矛盾リスクをfile:line…と共に出力します。
fhir-basics
nvidia
エージェントにFHIR R4 APIの動作方法、利用可能なリソース、検索パラメータを使ったクエリ方法、およびすべてのレスポンス形式を正しく解析する方法を教えます…
compileiq-validate-result
nvidia
検索が完了した後、かつスピードアップの申請やACFの発送の前に使用します。dump_results CSVを読み込み、トップK候補(単一目的)を抽出します…
changelog-audit
nvidia
リリース前にWarp CHANGELOG.mdを監査:失われたエントリを復元、ユーザー影響で並べ替え、エントリの文言を洗練、行折り返し、および(リリースブランチモードで)比較をバンプ…
maintain-dynamic-plugins
nvidia
NeMo Relayの動的プラグインローダー、マニフェスト、RustネイティブSDK、gRPCワーカープロトコル、PythonワーカーSDK、ドキュメント、テスト、およびリリースワークフローのカバレッジを維持する
dgx-diagnose
nvidia
一般的なDGX Station GB300の問題(CUDAクラッシュ、誤ったGPUターゲット、vLLM/SGLangコンテナのバグ、MIG状態の問題、NVLink/Fabric Managerエラーなど)を診断します。