Deepstream Profile Pipeline

DeepStream Profiling Skill

Profile-driven pipeline creation. When the user indicates they want an efficient DeepStream

pipeline, this skill replaces guesswork with two measured numbers — **inference plateau

batch and HW ceiling** — and derives every other config from them. Then it profiles the

E2E pipeline with Nsight Systems and reports per-plugin NVTX timings.

Model- and pipeline-agnostic. The skill assumes only that the inference element is

nvinfer or nvinferserver (so model dims, precision, and batch knobs are settable through

the standard config). It works for detection (with or without tracker), classification,

segmentation, VLM, and embedding pipelines. Source can be file, RTSP, USB camera, or any

mix. The skill reads the user's actual config to discover model dims / target FPS / source

properties — it does NOT assume any particular model, codec, or resolution.

> Constraint. Terminal only. Use nsys profile to capture and nsys stats to extract.

> Do not depend on Nsight Lens or any GUI.

When to trigger

Activate this skill at pipeline creation time when the user's ask carries efficiency

intent. Concrete triggers:

• "build an efficient / fast / performant / optimized pipeline"
• "give me a pipeline that runs well on this GPU"
• "benchmark / profile / measure / tune / optimize this pipeline"
• "I want to run N streams at M FPS"
• "how many streams can this GPU handle"
• user explicitly asks for nsys or Nsight

For plain "build a pipeline" / "display this video" / "save this stream" with no perf intent,

hand off to the deepstream-generate-pipeline skill instead.

The 6-stage flow

Run the stages in order. Stage 0 fires *before* the pipeline is generated, so the user

starts from a perf-tuned skeleton. Stages 1–5 measure and verify.

Stage 0 — Preset-apply (at pipeline-creation time)

Trigger: any time the coding agent is about to generate a new DS pipeline AND the user's

prompt carries efficiency intent (see "When to trigger" above).

Action: pre-apply these defaults *without prompting*. The user does not need to know any of

them; they just get a pipeline that's already in the right shape.

| Knob | Default value | Skip when |

|---|---|---|

| nvinfer.network-mode | 1 (INT8) if a calibration file is present at int8-calib-file=<path>, else 2 (FP16). Never FP32. | Model has no INT8 calibration AND the user explicitly says "FP32". |

| nvinfer.model-engine-file | Pre-built .engine path | Always set. Force a one-shot prebuild before measurement. |

| nvinfer.infer-dims | 3;<H>;<W> matching the model's native input | Always set, even for static-shape ONNX (harmless). |

| nvstreammux.batch-size | min(N_streams, 16) until microbench refines it | — |

| nvstreammux.width / height | model's native input dims (read from the nvinfer config's infer-dims=3;H;W) | User explicitly asks for native source resolution at the muxer. |

| nvstreammux.batched-push-timeout | 1e6 / source_fps µs (33333 for 30 fps) | — |

| nvstreammux.nvbuf-memory-type | 0 (NVMM) | — |

| Decoder num-extra-surfaces | min(batch_size, 5) | — |

| Decoder cudadec-memtype | 0 (NVMM) | — |

| Sink | fakesink sync=False for the benchmark variant | User asked for on-screen display or on-disk recording (then keep OSD/tiler/encoder/sink and produce TWO variants). |

| OSD + tiler | omit | User asked for visible output. |

| Tracker ll-config-file | config_tracker_NvDCF_max_perf.yml (perf-tuned NvDCF preset shipped with DS 9.0) | Tracker not present. |

| Tracker tracker-width / height | 480 / 288 | — |

| Tracker enable-batch-process (in linked YAML) | 1 | — |

| Queue between source and pgie | max-size-buffers = batch_size × 4 | No queue requested (rare). |

| Kafka/message queue | max-size-buffers=2, leaky=2 | No Kafka. |

| Decode-side PerfMonitor | attach (in addition to pgie-side) | Pipeline is nvurisrcbin → pgie direct without intermediate queue. |

Why Stage 0 exists: without it, every newly generated pipeline starts from

display-first defaults and Stages 1–5 spend cycles fixing avoidable issues. Stage 0 is the

"don't write a bad pipeline in the first place" gate.

The student / API user never sees these knobs. The skill's response back to the user is in

plain English (FPS, stream count, observed bottleneck), not knob names.

The verification flow (Stages 1–5)

Run the stages in order. Do not skip a stage — later stages depend on earlier ones' outputs.

Stage 1 — NVTX coverage check

DeepStream plugins emit NVTX ranges natively; custom plugins and plain GStreamer-core

elements (queue, tee, h264parse, etc.) do not. Before profiling, list the elements the

pipeline uses and classify each.

• Read the pipeline definition (gst-launch string or pipeline.py).
• For each element, look it up in references/nvtx-coverage.md.
• Classify COVERED (emits NVTX in this DS / image / nsys combo) or UNINSTRUMENTED.
• MVP rule: the skill *prefers* per-plugin NVTX as confirmation but does not require it.

Decode-bound diagnosis works from microbench shape + nvidia-smi dmon; compute-bound from

CUDA kernel mix; memcpy from cuda_gpu_mem_time_sum. NVTX is a bonus.

• For UNINSTRUMENTED elements, the skill reports "not directly measurable in this build" and

still applies the closed-form R1–R6 knobs (which are derived from inputs, not from

per-plugin profile data).

• Auto-injecting NVTX for uninstrumented elements is out of scope for this version —

flag it as follow-up in the final report.

Output of Stage 1: a short coverage table, e.g.

nvurisrcbin       COVERED
nvstreammux       COVERED
nvinfer           COVERED
nvtracker         COVERED
queue_src         UNINSTRUMENTED — not re-tuned
fakesink          UNINSTRUMENTED — not re-tuned

Stage 2 — HW discovery

Run nvidia-smi and derive theoretical ceilings for the host GPU. Minimum queries:

# Identity + memory + compute
nvidia-smi --query-gpu=name,compute_cap,memory.total,memory.free,\
clocks.max.sm,clocks.max.memory,utilization.gpu \
--format=csv,noheader,nounits

# NVDEC / NVENC utilization (per-engine)
nvidia-smi --query-gpu=utilization.decoder,utilization.encoder \
--format=csv,noheader,nounits

# PCIe link width/gen (for H2D memcpy ceiling)
nvidia-smi --query-gpu=pcie.link.gen.current,pcie.link.width.current \
--format=csv,noheader,nounits

Derive from those numbers:

• Decode ceiling (fps): NVDEC_count × per-unit H265/H264 fps for the source resolution

(table in references/hw-ceiling-formulas.md).

• Compute ceiling (TOPS): SM count × clock × ops-per-clock at the target precision. Gives

an upper bound — real models hit 30–60% of this.

• Memory-bandwidth ceiling (GB/s): memory clock × bus width. Model weight reads +

activations should fit well under this.

• Memcpy ceiling (GB/s): PCIe gen × width × 0.8 practical. Only relevant if NVMM is

broken and H2D/D2H transfers appear in Stage 5.

Store the derived ceilings — they drive the Stage 5 "actual vs. theoretical" section.

Full formulas and the per-codec NVDEC throughput table:

references/hw-ceiling-formulas.md.

Stage 3 — Inference-only micro-benchmark

Run only the inference stage (source → streammux → nvinfer → fakesink), sweeping

batch-size to find the plateau. This isolates the model's true peak FPS from everything

else, and answers "how many streams fit into a single batch without FPS dropping?".

Sweep: batch-size ∈ {1, 2, 4, 8, 16, 32} (cap at N_streams and at GPU memory).

For each batch size:

• Set nvstreammux.batch-size = nvinfer.batch-size = B.
• Set nvstreammux.width/height = the model's native infer-dims (read from the nvinfer config).
• fakesink sync=False as the only branch.
• Run 30 s; measure FPS from measure_fps_probe (console) or DS PerfMonitor.
• Record (B, fps).

Plateau batch = the smallest B where increasing to 2×B yields < 5% FPS gain. That is the

target batch for the full pipeline.

If the user's N_streams ≤ plateau batch, set final batch = N_streams. Otherwise set final

batch = plateau batch and note that the pipeline will process streams in multiple batches

per tick.

Stage 4 — Derive configs

From (plateau_batch, HW_ceilings, N_streams, source_res, source_fps), set every tunable

knob at once. Do not tune one knob at a time — the derivation rules are closed-form.

Knobs to set, in order:

• Streammux: batch-size = final_batch, width/height = min(source_res, infer_dims),

batched-push-timeout = 1e6 / source_fps µs, nvbuf-memory-type = 0.

• Inference: batch-size = final_batch, `network-mode = 1 (INT8) if calib file exists

else 2 (FP16), interval = 0, infer-dims = model's native dims, model-engine-file =

pre-built .engine path`.

• Decoder (on nvurisrcbin / nvmultiurisrcbin / nvv4l2decoder):

num-extra-surfaces = min(final_batch, 5), cudadec-memtype = 0, nvbuf-memory-type = 0.

• Tracker (if present): enable-batch-process = 1, tracker res 480×288, point

ll-config-file at config_tracker_NvDCF_max_perf.yml.

• Queues (if present between decoder and streammux, or streammux and nvinfer):

max-size-buffers = final_batch × 2. Kafka/message branches: `leaky=2,

max-size-buffers=2`.

Full derivation table with each formula and a one-line "why":

references/config-derivation-rules.md.

Write the derived values into the user's config files (pgie_config.yml,

tracker_config.yml, pipeline.py source properties, any deepstream-app .txt). Always

Read before Edit. Keep edits surgical — do not reformat unrelated lines.

Stage 5 — E2E profile + report

Run the E2E pipeline under nsys profile and extract per-plugin timings via nsys stats.

Capture:

TS=$(date +%Y%m%d_%H%M%S)
OUT=/tmp/ds_profile_${TS}
nsys profile \
  --trace=cuda,nvtx,osrt \
  --gpu-metrics-devices=all \
  --cuda-memory-usage=true \
  --force-overwrite=true \
  --duration=30 \
  --output=${OUT} \
  <your-pipeline-launch-command>

Extract:

# Per-kernel GPU time (top 10)
nsys stats --report cuda_gpu_kern_sum --format csv ${OUT}.nsys-rep | head -20

# Per-NVTX-range time (top 10) — this is the DS per-plugin breakdown
nsys stats --report nvtx_sum --format csv ${OUT}.nsys-rep | head -20

# Memcpy totals
nsys stats --report cuda_gpu_mem_time_sum --format csv ${OUT}.nsys-rep

# GPU metrics (SM activity, DRAM throughput) — requires --gpu-metrics-devices
nsys stats --report gpu_metric_gpu_util_sum --format csv ${OUT}.nsys-rep

Full command reference: references/nsys-cli-recipes.md.

Report (Markdown, to stdout — no external UI):

## Profile summary

**Hardware**: <name>, <mem_total> GB, SM x<sm>, NVDEC x<nvdec>, PCIe Gen<g> x<w>
**Ceilings**: decode <X> fps, compute ~<Y> TOPS @ INT8, memory <Z> GB/s

**Inference plateau**: batch=<B>, peak=<F> fps per batch → <F × B> fps aggregate

**E2E measured**: <actual> fps  (=<pct>% of inference plateau)

### Per-plugin time (from NVTX) — only for plugins emitting NVTX in this build

| Plugin          | Share of wall time | GPU / CPU | Notes |
|-----------------|--------------------|-----------|-------|
| nvinfer         | <pct>%             | GPU       | (always emitted; if absent, NVTX injection is broken) |
| nvdsosd         | <pct>%             | GPU       | (when in pipeline) |
| ...             | ...                | ...       | (other plugins as the verification probe shows) |

(Numbers above are illustrative — fill in from `nsys stats --report nvtx_sum`. Plugins
that don't emit NVTX in your DS / image combo simply don't appear; that's not a bug, it's
the limit of what NVTX captures here. See `references/nvtx-coverage.md`.)

### Applied configs (sample shape; values come from R1–R6 + Stage 3 measurements)

- `nvstreammux.batch-size = <plateau_batch>`
- `nvinfer.network-mode = 1 (INT8)` if calibration available, else `2 (FP16)`
- decoder `num-extra-surfaces = min(plateau_batch, 5)`
- queue between source and pgie, `max-size-buffers = plateau_batch × 4`
- ... (full list per the user's pipeline shape)

### Uninstrumented (skipped re-tune)

List the elements that didn't emit NVTX in this build (typically the closed-source binary
plugins — see `references/nvtx-coverage.md`) plus plain GStreamer-core helpers. Report
them so the user knows what wasn't directly measurable.

Keep the summary terse. Raw nsys stats CSV goes into the temp file, not the response.

Reference documents

| Document | Use when |

|----------|----------|

| references/nvtx-coverage.md | Stage 1 — classifying each pipeline element as COVERED or UNINSTRUMENTED. |

| references/hw-ceiling-formulas.md | Stage 2 — turning nvidia-smi output into decode / compute / memory ceilings. |

| references/config-derivation-rules.md | Stage 4 — per-knob formula keyed to (plateau_batch, HW, N_streams, source_res, source_fps). |

| references/nsys-cli-recipes.md | Stages 3 & 5 — exact nsys profile / nsys stats invocations. |

Non-goals (this version)

• No Nsight Lens / no GUI. Terminal only.
• No NVTX auto-injection for uninstrumented plugins. MVP skips their knobs. Future work.
• No iterative tune-measure-tune loop. Stage 4 derives configs once from closed-form

rules; Stage 5 measures and reports. If the user wants to keep tuning, they can re-invoke

the skill with updated inputs.

Related skills

• deepstream-generate-pipeline — upstream pipeline generation. This skill

assumes a pipeline already exists or is about to be generated.

• deepstream-byovm — HF → TensorRT engine building. Run first if the user

brought a new model; come here after.

Notes

• Lives in skills/deepstream-profile-pipeline/ alongside the other DS skills, per

the repo convention in CLAUDE.md.

• For ground-truth on any plugin's properties (types, defaults, ranges) and pad caps,

query the loaded binary inside the DS container:

  gst-inspect-1.0 nvinfer
  gst-inspect-1.0 nvstreammux
  gst-inspect-1.0 nvurisrcbin   # works on closed-source binary plugins too
  gst-inspect-1.0 | grep ^nv    # list every NVIDIA-specific element this build ships

Plugin naming convention: any element prefixed nv* is NVIDIA DeepStream-specific

(NVMM-capable, may emit NVTX); everything else is upstream GStreamer-core (no NVMM,

never emits DS NVTX). Use this prefix as the first-pass classifier when triaging an

unfamiliar pipeline.

• The open-source subset of plugin code lives under

/opt/nvidia/deepstream/deepstream/sources/gst-plugins/ if you need to read the

implementation (only some plugins are open — closed ones must be inspected via

gst-inspect-1.0 and behaviour observed at runtime).

<!-- Signing refresh marker. -->