Benchmark Performance#
This document provides instructions on how to run performance benchmarks for the Vision AI applications using the provided benchmarking scripts. The script determines the maximum number of concurrent video streams a system can process (stream density) while maintaining a target performance level.
Prerequisites#
The
metro-ai-suiterepository must be cloned to your system.curl,jq,gawk,ffmpeg, andbcutilities installedDocker with the
docker composeplugin installed.The application must be installed with
./install.sh <app-name>(e.g.,./install.sh loitering-detection). Do not rundocker compose up; the benchmark script starts and stops the application stack itself.
Step 1: Understand the Benchmarking Script#
The core of the benchmarking process is the calc_stream_density.sh script, located in the
metro-vision-ai-app-recipe/ directory. This script automates the process of starting video
streams, monitoring their performance (Frames Per Second - FPS), and calculating key
performance indicators (KPIs) to find the maximum sustainable stream density.
Fresh Application Stack per Round#
The script manages the application lifecycle itself, so each benchmark round starts from a clean state:
Before benchmarking, it runs
docker compose down -vto clear any running stack.At the start of every round (every stream count tested), it runs
docker compose up -dand waits until the DL Streamer Pipeline Server is ready (timeout set by theCOMPOSE_READY_TIMEOUTenvironment variable, default 300 seconds).At the end of every round, it stops the pipelines and runs
docker compose down -v --remove-orphans.If the script is interrupted (Ctrl+C), the stack is also brought down.
Stream Density Logic#
The script determines the maximum stream density automatically: you supply only the pipeline
name and the FPS target, and the script discovers the search range itself using an exponential
ramp-up followed by a bisect. Here is a summary of the logic from the
calc_stream_density.sh script:
Initialization: The script starts at 1 stream. It tracks the highest stream count that met the target (
lo, also reported astns) and the lowest stream count that failed it (hi). No user-provided bounds are needed.Phase 1 - Exponential ramp-up: The script measures 1, 2, 4, 8, 16, … streams. In each iteration:
It runs a workload with the current number of streams (
ns).It measures the
throughput min(the lowest FPS achieved among all streams) and compares it to thetarget_fps.If the target is met, the stream count is doubled and the search continues.
The first stream count that fails the target becomes the upper edge (
hi) and ends this phase.
Phase 2 - Bisect: The script binary-searches between the last passing count (
lo) and the first failing count (hi), testing the midpoint each time and narrowing the interval, until the two are adjacent.Convergence: The final value of
lois the highest number of streams that met the performance target, reported as the final stream density. If even a single stream fails the target, the script reports that the target is not achievable. A safety ceiling (default 64 streams, override with theMAX_STREAMSenvironment variable) bounds the ramp-up on very capable hardware.
Average FPS Calculation#
During each test run, the script logs the avg_fps for every active pipeline instance at
regular intervals. At the end of the run, an awk script processes these logs to calculate
several KPIs for the collection of FPS samples from each stream:
Percentile Throughput: Calculates a specific percentile (e.g., 90th) of the FPS values to ignore outliers.
Average Throughput: The mean FPS across all streams.
Median Throughput: The median FPS value.
Cumulative Throughput: The sum of the FPS from all streams.
Min Throughput: The lowest (worst-case) FPS achieved among all streams. This value is critical for the stream density calculation.
NStreams Mode#
The script also supports an NStreams mode for testing fixed stream counts without binary search. This mode allows you to run multiple pipeline types in parallel simultaneously with predefined stream counts for each pipeline.
When to use NStreams Mode:
When you want to test specific stream count combinations across different pipelines (e.g., CPU and GPU simultaneously).
To measure combined workload performance of heterogeneous pipelines on the same system.
When you already know the desired stream counts and want to verify their performance.
How NStreams Mode Works:
The script starts all specified pipelines with their respective fixed stream counts concurrently.
Monitoring continues for the specified duration (default 60 seconds).
KPIs are computed for the combined workload across all pipelines.
No binary search is performed; results are based on the exact stream counts provided.
Recommended Pipeline Parameters#
These are the recommended parameters by Edge Benchmarking and Workloads team for workload with similar characteristics. These are configurable parameters that can be adjusted based on your specific requirements:
inference-region=0 inference-interval=3 batch-size=8 nireq=2 ie-config="GPU_THROUGHPUT_STREAMS=2" threshold=0.7
Parameter Descriptions:
inference-region=0: Use full frame.inference-interval=3: Run inference on every 3rd frame.batch-size=8: Process 8 frames in a single batch for better GPU utilization.nireq=2: Number of inference requests to run in parallel.ie-config="GPU_THROUGHPUT_STREAMS=2": OpenVINO™ engine streams configuration.threshold=0.7: Detection confidence threshold (70%).
Step 2: Prepare for Benchmarking#
Navigate to Script Directory: Open a terminal and navigate to the
metro-vision-ai-app-recipedirectory.cd metro-ai-suite/metro-vision-ai-app-recipe/
Install the Application: Follow the
get-started.mdguide for the application (e.g., Get Started for Loitering Detection) only up to and including the./install.shstep. This downloads the models and videos and generates the.envanddocker-compose.ymlfiles used by the benchmark../install.sh loitering-detectionNote
Do not run
docker compose upor./sample_start.sh. Thecalc_stream_density.shscript runsdocker compose upanddocker compose downfor every benchmark round. Any stack that is already running is brought down when the benchmark starts.
Step 3: Run the Benchmark#
Note
The default parameters are set based on best know methods recommended by Edge Workloads and Benchamarks group for workload with similar characteristics. These parameters can be modified when starting the pipelines.
The calc_stream_density.sh script requires only a pipeline name; the stream count range is
determined automatically. The available pipelines are defined in the
benchmark_app_payload.json file located within each application’s directory (e.g.,
loitering-detection/).
Example Payload with Detection and Classification
The benchmark_app_payload.json file contains an array of pipeline configurations. Each
configuration specifies the pipeline name and a payload with parameters for source, destination,
and AI models. The script uses the pipeline name to select the corresponding payload for
benchmarking.
Here is an example of a GPU pipeline configuration that includes both detection-properties
and classification-properties with additional parameters:
{
"pipeline": "object_tracking_gpu",
"payload": {
"source": {
"uri": "file:///home/pipeline-server/videos/VIRAT_S_000101_looped.mp4",
"type": "uri"
},
"destination": {
"metadata": {
"type": "mqtt",
"topic": "object_detection_$x",
"publish_frame": false
},
"frame": {
"type": "webrtc",
"peer-id": "object_detection_$x",
"overlay-properties": {
"font-scale": 1.0,
"draw-txt-bg": false
}
}
},
"parameters": {
"detection-properties": {
"model": "/home/pipeline-server/models/intel/pedestrian-and-vehicle-detector-adas-0001/FP16/pedestrian-and-vehicle-detector-adas-0001.xml",
"device": "GPU",
"inference-interval": 3,
"inference-region": 0,
"batch-size": 8,
"nireq": 2,
"ie-config": "GPU_THROUGHPUT_STREAMS=2",
"pre-process-backend": "va-surface-sharing",
"threshold": 0.7
}
}
}
}
Example: Running Stream Density Benchmark for Loitering Detection#
This example will find the maximum number of loitering detection streams that can run on the CPU while maintaining at least 15 FPS.
Execute the
calc_stream_density.shscript, providing the desired pipeline name (object_tracking_cpuin this case). The script finds the stream count on its own.# Usage: ./calc_stream_density.sh -p <pipeline_name> -t <target_fps> ./calc_stream_density.sh -p object_tracking_cpu -t 15
The script will output its progress as it tests different stream counts. The final output will show the optimal stream density found.
✅ FINAL RESULT: Stream-Density Benchmark Completed! stream density: 8 ====================================================== KPIs for the optimal configuration (8 streams): throughput #1: 29.98 throughput #2: 29.98 ... throughput #8: 29.98 throughput median: 29.98 throughput average: 29.98 throughput stdev: 0 throughput cumulative: 239.84 throughput min: 29.98
Example: Running Multiple Pipelines with Fixed Stream Counts#
To test multiple pipelines in parallel with predefined stream counts (NStreams mode), use the -nstreams flag. This example runs 8 GPU streams and 6 NPU streams for loitering detection concurrently:
# Usage: ./calc_stream_density.sh -p <pipeline1> <pipeline2> ... -nstreams <count1> <count2> ...
./calc_stream_density.sh -p object_tracking_gpu object_tracking_npu -nstreams 8 6 -t 15 -i 60
Parameters:
-p object_tracking_gpu object_tracking_npu: Two pipeline names to run in parallel.-nstreams 8 6: 8 streams for object_tracking_gpu, 6 streams for object_tracking_npu (order must match pipeline order).-t 15: Target FPS threshold (optional, default 14.95).-i 60: Monitoring duration in seconds (optional, default 60).
The script will start all specified pipelines and monitor their combined performance. Final output shows aggregated KPIs:
✅ FINAL RESULT: Nstreams-mode Pipeline Run Completed!
Pipelines : object_tracking_gpu object_tracking_npu
Streams : 8 6
Total : 14 streams
======================================================
KPIs (all 14 streams combined):
throughput median: 28.5
throughput average: 28.8
...
throughput min: 27.2
Step 4: Stop the Benchmark#
The script brings the application stack down (docker compose down) after every round,
including the last one, and also when it is interrupted with Ctrl+C. No manual cleanup is
needed. If the script was terminated abnormally (e.g., kill -9), bring the stack down
manually:
docker compose down -v --remove-orphans
Intel, the Intel logo, OpenVINO, and the OpenVINO logo are trademarks of Intel Corporation or its subsidiaries.