Optimizer

DLS Optimizer is a tool for helping users discover more optimal versions of the pipelines they run on DL Streamer. It will explore different modifications to your pipeline that are known to increase performance and measure them. As a result, you can expect to receive a pipeline that is better suited to your setup.

Optimizations involve modifying inference elements that are part of DL Streamer, as well as searching for pre- and post-processing elements better suited for the pipeline.

Modification of inference elements currently covers:

• Discovering more suitable devices

• Adjusting the batching of frames in an element

• Adjusting the number of inference requests done simultaneously (nireqs)

Limitations

Currently the DLS Optimizer focuses mainly on DL Streamer elements, specifically the gvadetect and gvaclassify. The produced pipeline could still have potential for further optimization by transforming other elements.

Multi-stream pipelines (those utilizing the tee element) are also currently not supported.

Prerequisites

Before using the DLS Optimizer, ensure you have:

• Installed DL Streamer intallation steps.

• Installed any necessary DL Streamer python dependencies.

• Configured environment variables for the current terminal session.

• Installed the OpenVINO python library.

   python3 -m venv ~/python3venv
   source ~/python3venv/bin/activate
   source /opt/intel/dlstreamer/scripts/setup_dls_env.sh
   cd /opt/intel/dlstreamer/scripts/optimizer
   pip install openvino==2026.2

Using the optimizer as a tool

Note
This example assumes your working directory is the optimizer directory /opt/intel/dlstreamer/scripts/optimizer

python3 . MODE [OPT] -- PIPELINE

Arguments:
  MODE      The type of optimization that will be performed on the pipeline.
            Possible values are "fps" and "power".

            fps - the optimizer will explore possible alternatives
                  for the pipeline, trying to locate versions that
                  have increased performance measured by fps.

            power - the optimizer will explore possible alternatives
                    for the pipeline, trying to locate versions that
                    consume the least amount of watts.

  PIPELINE  A string representing a pipeline in the GStreamer notation
            which the tool will attempt to optimize.

Options:
    --search-duration SEARCH_DURATION   How long should the optimizer search for better pipelines.
    --sample-duration SAMPLE_DURATION   How long should every pipeline be sampled for performance.
    --detection-threshold THRESHOLD     Minimum threshold of detections that tested pipelines are
                                        not allowed to cross in order to count as valid alternatives.
    --fps-limit LIMIT                   Minimum fps that every valid pipeline must achieve.
    --power-limit LIMIT                 Maximum power (watts) that every valid pipeline cannot cross.
    --power-metrics-endpoint URL        URL leading to the Prometheus endpoint of a Metrics Manager.
                                        Required for power-based optimization.
    --enable-cross-stream-batching      Enable cross stream batching for inference elements in fps mode.
    --maximize-streams                  When optimizing, try to pack as many parallel streams as
                                        possible without crossing the fps or power limit.
                                        Requires at least one of --fps-limit or --power-limit.
    --allowed-devices ALLOWED_DEVICES   List of allowed devices (CPU, GPU, NPU) to be used by the optimizer.
                                        If not specified, all available, detected devices will be used.
                                        Tool does not support discrete GPU selection.
                                        eg.--allowed-devices CPU NPU,--allowed-devices GPU
    --batch-sizes BATCH_SIZES [BATCH_SIZES ...]
                                        List of batch sizes to be considered by the optimizer.
    --nireq-sizes NIREQ_SIZES [NIREQ_SIZES ...]
                                        List of nireq sizes to be considered by the optimizer.
    --optimization-profile PROFILE       Configuration preset for optimization behavior.
                                        Possible values: coarse, fine, default.
    --log-level LEVEL                   Configure the logging detail level.
    -v, --verbose                       Print information about every candidate pipeline investigated during
                                        optimization process.
    -o, --output OUTPUT_FILE            Save optimization results to a file in JSON format.

search-duration default: 300 seconds
Increasing the search duration will increase the chances of discovering more performant pipelines.

sample-duration default: 10 seconds
Increasing the sample duration will improve the stability of the search.

fps-limit
Minimum fps that every valid pipeline must achieve. When optimizing for streams, increasing this limit will improve the performance of each individual stream, but the final result is liable to support less streams overall.

power-limit
Maximum power consumption (watts) that every valid pipeline cannot cross.

power-metrics-endpoint
URL leading to the Prometheus endpoint of a Metrics Manager instance. Required when using the power optimization mode.

enable-cross-stream-batching
Levy the inference instance feature of DL Streamer to batch work across multiple streams in fps mode.

maximize-streams
When enabled, the optimizer will try to pack as many parallel streams as possible without crossing either the fps-limit or power-limit. At least one of these limits must be configured.

allowed-devices
Allows you to limit the set of devices that will be considered during the optimization process.

batch-sizes default: 1 2 4 8 16 32
Defines candidate batch-size values that can be applied to supported inference elements while searching.

nireq-sizes default: 1 2 3 4 5 6 7 8
Defines candidate nireq values that can be applied to supported inference elements while searching.

optimization-profile default: default
Selects a preset that controls optimization aggressiveness and runtime.

• coarse - fast search with minimal exploration (uses short timings and disables batch-sizes/nireq-sizes exploration).

• fine - balanced search with focused exploration (batch-sizes: 1 4 8, nireq-sizes: 4 8).

• default - full search using user-provided values for search-duration, sample-duration, batch-sizes, and nireq-sizes.

Note
When --optimization-profile is coarse or fine, the tool uses profile-specific values for search configuration. In these profiles, user-provided --batch-sizes, --nireq-sizes, and timing flags are being overridden by the selected preset.

log-level default: INFO
Available log levels are: CRITICAL, FATAL, ERROR, WARN, INFO, DEBUG.

verbose
Prints extra information about the candidate pipelines which were examined during the optimization process.

output
Provide a path representing a file which will be used to save results information in JSON format.

Note
Search duration and sample duration both affect the amount of pipelines that will be explored during the search.
The total amount should be approximately search_duration / sample_duration pipelines.

Pausing and resuming

While the optimizer is running, you can pause and resume the search at any time by pressing the Spacebar in the terminal where the tool is running.

Pausing stops the current pipeline sample mid-run and suspends the search loop. The optimizer will not start any new pipeline test until it is resumed. Resuming (pressing Space again) continues to loop through the candidates starting from the next one. The remaining search duration is preserved accurately: time spent while paused is not counted against the budget.

Note
The pause key is only active while the optimizer is running as a CLI tool in terminal with keyboard input available. It is not available in non-interactive or piped environments.

---

Example

 python3 . fps -- urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink
[__main__] [    INFO] - GStreamer initialized successfully
[__main__] [    INFO] - GStreamer version: 1.26.6
[__main__] [    INFO] - Detected GPU Device
[__main__] [    INFO] - No NPU Device detected
[__main__] [    INFO] - Sampling for 10 seconds...
FpsCounter(last 1.00sec): total=46.87 fps, number-streams=1, per-stream=46.87 fps
FpsCounter(average 1.00sec): total=46.87 fps, number-streams=1, per-stream=46.87 fps
FpsCounter(last 1.01sec): total=43.70 fps, number-streams=1, per-stream=43.70 fps
FpsCounter(average 2.01sec): total=45.28 fps, number-streams=1, per-stream=45.28 fps

...

FpsCounter(last 1.09sec): total=73.45 fps, number-streams=1, per-stream=73.45 fps
FpsCounter(average 8.70sec): total=73.65 fps, number-streams=1, per-stream=73.65 fps
[__main__] [    INFO] - Best found pipeline: urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 !decodebin3!gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml device=GPU pre-process-backend=va-surface-sharing batch-size=2 nireq=2 ! queue ! gvawatermark ! fakesink with fps: 81.987923.2

In this case the optimizer started with a pipeline that ran at ~45fps, and found a pipeline that ran at ~82fps instead. The specific improvements were:

• replacing the decodebin with the decodebin3 element.

• configuring the gvadetect element to use GPU for processing

• setting the batch-size parameter to 2

• setting the nireq parameter to 2

Using the optimizer as a library

The easiest way of importing the optimizer into your scripts is to include it in your PYTHONPATH environment variable:
export PYTHONPATH=/opt/intel/dlstreamer/scripts/optimizer

Targets which are exported in order to facilitate usage inside of scripts:

preprocess_pipeline(pipeline) -> processed_pipeline

• pipeline: string - A string containing a valid DL Streamer pipeline.

• processed_pipeline: string - A string containing the pipeline with all relevant substitutions.

Perform quick search and replace for known combinations of elements with more performant alternatives.

---

DLSOptimizer class

Initialized without any arguments

optimizer = DLSOptimizer()

Result dictionary

Methods that measure or optimize pipelines return a result dictionary with the following keys:

Key	Type	Description	Present in
fps	float	Measured frames per second	All results
power	float | None	Measured power consumption in watts (average over sample)	All results (requires metrics endpoint)
streams	int	Number of concurrent streams tested	Stream-optimization results only

Methods

get_baseline_pipeline() -> pipeline, result

• pipeline: string - The baseline pipeline from which optimization started.

• result: dict - Result dictionary (see above).

Returns information about the original pipeline used in the optimization process. Returned values are meaningless until at least one optimization operation is performed.

optimizer = DLSOptimizer()
for (_, _) in optimizer.iter_optimize_for_fps(pipeline):
    pass
pipeline, result = optimizer.get_baseline_pipeline()
print(f"Baseline FPS: {result['fps']}")

---

get_optimal_pipeline() -> pipeline, result

• pipeline: string - The best pipeline found during optimization.

• result: dict - Result dictionary (see above).

Returns information about the best pipeline found during the optimization process. Returned values are meaningless until at least one optimization operation is performed.

optimizer = DLSOptimizer()
for (_, _) in optimizer.iter_optimize_for_fps(pipeline):
    pass
best_pipeline, result = optimizer.get_optimal_pipeline()
print(f"Best FPS: {result['fps']}, Streams: {result['streams']}")

---

set_sample_duration(duration)

• duration: int - The duration of sampling each candidate pipeline in seconds, default 10.

Configures the sample duration used in optimization sessions.

optimizer = DLSOptimizer()
optimizer.set_sample_duration(15)

---

set_detections_error_threshold(threshold)

• threshold: float - The threshold of counted detections, between 0.0 and 1.0, default 0.95.

Minimum threshold of detections that tested pipelines are not allowed to cross in order to count as valid alternatives.

optimizer = DLSOptimizer()
optimizer.set_detections_error_threshold(0.8)

---

enable_cross_stream_batching(enable)

• enable: bool - Enable the cross stream batching feature, default False.

Levy the inference instance feature of DL Streamer to batch work across multiple streams when optimizing for fps.

optimizer = DLSOptimizer()
optimizer.enable_cross_stream_batching(True)

---

set_maximize_streams(maximize)

• maximize: bool - Enable stream maximization, default False.

When enabled, the optimizer will try to pack as many parallel streams as possible without crossing the configured fps or power limit. At least one limit (set_fps_limit or set_power_limit) must be configured. This replaces the deprecated optimize_for_streams and iter_optimize_for_streams methods.

optimizer = DLSOptimizer()
optimizer.set_maximize_streams(True)
optimizer.set_fps_limit(30)
optimized_pipeline, result = optimizer.optimize_for_fps(pipeline)
print(f"Streams: {result['streams']}, FPS: {result['fps']}")

---

set_fps_limit(limit)

• limit: float - The minimum fps that every valid pipeline must achieve.

Configures the minimum fps limit. Pipelines accepted during optimization are not allowed to fall below this threshold.

optimizer = DLSOptimizer()
optimizer.set_fps_limit(45)

---

set_power_limit(limit)

• limit: float - The maximum power consumption (watts) that every valid pipeline cannot cross.

Configures the maximum power limit for valid pipelines.

optimizer = DLSOptimizer()
optimizer.set_power_limit(15.0)

---

set_metrics_url(url)

• url: string - URL leading to the Prometheus endpoint of a Metrics Manager.

Configures the power metrics endpoint used for measuring power consumption. Required for power-based optimization.

optimizer = DLSOptimizer()
optimizer.set_metrics_url("http://localhost:9090/")

---

set_batch_sizes(sizes)

• sizes: list[int] - A list of batch size values to consider during optimization, default [1, 2, 4, 8, 16, 32].

Defines candidate batch-size values that can be applied to supported inference elements while searching.

optimizer = DLSOptimizer()
optimizer.set_batch_sizes([1, 4, 8])

---

set_nireq_sizes(sizes)

• sizes: list[int] - A list of nireq values to consider during optimization, default [1, 2, 3, 4, 5, 6, 7, 8].

Defines candidate nireq values that can be applied to supported inference elements while searching.

optimizer = DLSOptimizer()
optimizer.set_nireq_sizes([2, 4, 6, 8])

---

set_allowed_devices(devices)

• devices: list[string] - A list of device identifiers.

Limits the set of devices which will be considered during the optimization process.

optimizer = DLSOptimizer()
optimizer.set_allowed_devices(["CPU", "GPU"])

---

optimize_for_fps(pipeline, search_duration) -> optimized_pipeline, result

• pipeline: string - A string containing a valid DL Streamer pipeline.

• search_duration: int - The duration of searching for better pipelines, default 300.

• optimized_pipeline: string - A string containing the best performing pipeline that has been found during the search.

• result: dict - Result dictionary (see above).

Runs a series of optimization steps on the pipeline searching for a version with better performance measured by fps.

pipeline = "urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink"
optimizer = DLSOptimizer()
optimized_pipeline, result = optimizer.optimize_for_fps(pipeline)
print(f"Best pipeline: {optimized_pipeline} @ {result['fps']} fps")

---

iter_optimize_for_fps(pipeline) -> optimized_pipeline, result

• pipeline: string - A string containing a valid DL Streamer pipeline.

• optimized_pipeline: string - A string containing a candidate pipeline that has been tested.

• result: dict | None - Result dictionary (see above), or None if the candidate failed validation.

Runs a series of optimization steps on the pipeline searching for version with better performance measured by fps. Returns each and every candidate pipeline that has been considered.

pipeline = "urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink"
optimizer = DLSOptimizer()
for (pipeline, result) in optimizer.iter_optimize_for_fps(pipeline):
    if result:
        print(f"Tested: {pipeline} @ {result['fps']}")
    else:
        print(f"Failed: {pipeline}")
best_pipeline, best_result = optimizer.get_optimal_pipeline()
print(f"Optimal pipeline: {best_pipeline} @ {best_result['fps']}")

---

optimize_for_power(pipeline, search_duration) -> optimized_pipeline, result

• pipeline: string - A string containing a valid DL Streamer pipeline.

• search_duration: int - The duration of searching for better pipelines, default 300.

• optimized_pipeline: string - A string containing the best performing pipeline that has been found during the search.

• result: dict - Result dictionary (see above).

Runs a series of optimization steps on the pipeline searching for a version with the lowest power consumption. Requires a metrics endpoint to be configured via set_metrics_url.

pipeline = "urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink"
optimizer = DLSOptimizer()
optimizer.set_metrics_url("http://localhost:9090/")
optimized_pipeline, result = optimizer.optimize_for_power(pipeline)
print(f"Best pipeline: {optimized_pipeline} @ {result['power']} watts")

---

iter_optimize_for_power(pipeline) -> candidate_pipeline, result

• pipeline: string - A string containing a valid DL Streamer pipeline.

• candidate_pipeline: string - A string containing a candidate pipeline that has been tested.

• result: dict | None - Result dictionary (see above), or None if the candidate failed validation.

Runs a series of optimization steps on the pipeline searching for a version with the lowest power consumption. Returns each and every candidate pipeline that has been considered. Requires a metrics endpoint to be configured via set_metrics_url.

pipeline = "urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink"
optimizer = DLSOptimizer()
optimizer.set_metrics_url("http://localhost:9090/")
for (pipeline, result) in optimizer.iter_optimize_for_power(pipeline):
    if result:
        print(f"Tested: {pipeline} @ {result['power']} watts")
    else:
        print(f"Failed: {pipeline}")
best_pipeline, best_result = optimizer.get_optimal_pipeline()
print(f"Optimal pipeline: {best_pipeline} @ {best_result['power']} watts")

---

Example:

from optimizer import DLSOptimizer

pipeline = "urisourcebin buffer-size=4096 uri=https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 ! decodebin ! gvadetect model=/home/optimizer/models/public/yolo11s/INT8/yolo11s.xml ! queue ! gvawatermark ! fakesink"

optimizer = DLSOptimizer()
optimizer.set_sample_duration(15)
optimized_pipeline, result = optimizer.optimize_for_fps(pipeline, search_duration = 600)
print("Best discovered pipeline: " + optimized_pipeline)
print("Measured fps: " + str(result["fps"]))

Controling the measurement

The point at which performance is being measured can be controlled by pre-emptively inserting a gvafpscounter element into your pipeline definition. For pipelines which lack such an element, the measurement is done after the last inference element supported by the optimizer tool.