Running on multiple devices simultaneously

Introducing the Multi-Device Plugin (C++)

C++

The Multi-Device plugin automatically assigns inference requests to available computational devices to execute the requests in parallel. By contrast, the Heterogeneous plugin can run different layers on different devices but not in parallel. The potential gains with the Multi-Device plugin are:

  • Improved throughput from using multiple devices (compared to single-device execution)

  • More consistent performance, since the devices share the inference burden (if one device is too busy, another can take more of the load)

Note that with Multi-Device the application logic is left unchanged, so you don’t need to explicitly compile the model on every device, create and balance the inference requests and so on. From the application point of view, this is just another device that handles the actual machinery. The only thing that is required to leverage performance is to provide the multi-device (and hence the underlying devices) with enough inference requests to process. For example, if you were processing 4 cameras on the CPU (with 4 inference requests), it might be desirable to process more cameras (with more requests in flight) to keep CPU and GPU busy via Multi-Device.

The setup of Multi-Device can be described in three major steps:

  1. Prepare configure for each device.

  2. Compile the model on the Multi-Device plugin created on top of a (prioritized) list of the configured devices with the configure prepared in step one.

  3. As with any other CompiledModel call (resulting from compile_model), you create as many requests as needed to saturate the devices.

These steps are covered below in detail.

Defining and Configuring the Multi-Device Plugin

Following the OpenVINO™ convention of labeling devices, the Multi-Device plugin uses the name “MULTI”. The only configuration option for the Multi-Device plugin is a prioritized list of devices to use:

Parameter name

Parameter values

Default

Description

ov::device::priorities

comma-separated device names with no spaces

N/A

Prioritized list of devices

You can set the priorities directly as a string.

Basically, there are three ways to specify the devices to be use by the “MULTI”:

ov::Core core;
std::shared_ptr<ov::Model> model = core.read_model("sample.xml");
// the "MULTI" device is (globally) pre-configured with the explicit option
core.set_property("MULTI", ov::device::priorities("HDDL,GPU"));
ov::CompiledModel compileModel0 = core.compile_model(model, "MULTI");

// configuration of the "MULTI" is part of the compile configuration (and hence specific to the model):
ov::CompiledModel compileModel1 = core.compile_model(model, "MULTI", ov::device::priorities("HDDL,GPU"));

// same as previous, but configuration of the "MULTI" is part
// of the name (so config is empty), also model-specific:
ov::CompiledModel compileModel2 = core.compile_model(model, "MULTI:HDDL,GPU");

Notice that the priorities of the devices can be changed in real time for the compiled model:

ov::Core core;
std::shared_ptr<ov::Model> model = core.read_model("sample.xml");
ov::CompiledModel compileModel = core.compile_model(model, "MULTI:HDDL,GPU");
// reverse the order of priorities
compileModel.set_property(ov::device::priorities("GPU,HDDL"));
// you can even exclude some device (HDDL here)
compileModel.set_property(ov::device::priorities("GPU"));
// and then return it back
compileModel.set_property(ov::device::priorities("GPU,HDDL"));
// but you cannot add new devices on the fly, 
// the next line will trigger the following exception:
// [ ERROR ] [NOT_FOUND] You can only change device
// priorities but not add new devices with the model's
// ov::device::priorities. CPU device was not in the original device list!
compileModel.set_property(ov::device::priorities("CPU,GPU,HDDL"));

Finally, there is a way to specify number of requests that the Multi-Device will internally keep for each device. Suppose your original app was running 4 cameras with 4 inference requests. You would probably want to share these 4 requests between 2 devices used in MULTI. The easiest way is to specify a number of requests for each device using parentheses: “MULTI:CPU(2),GPU(2)” and use the same 4 requests in your app. However, such an explicit configuration is not performance-portable and hence not recommended. Instead, the better way is to configure the individual devices and query the resulting number of requests to be used at the application level (see Configuring the Individual Devices and Creating the Multi-Device On Top).

Enumerating Available Devices

The OpenVINO Runtime API features a dedicated methods to enumerate devices and their capabilities. See the Hello Query Device C++ Sample. This is example output from the sample (truncated to device names only):

./hello_query_device
Available devices:
    Device: CPU
...
    Device: GPU.0
...
    Device: GPU.1
...
    Device: HDDL

A simple programmatic way to enumerate the devices and use with the multi-device is as follows:

ov::Core core;
std::shared_ptr<ov::Model> model = core.read_model("sample.xml");
std::vector<std::string> availableDevices = core.get_available_devices();
std::string all_devices;
for (auto && device : availableDevices) {
    all_devices += device;
    all_devices += ((device == availableDevices[availableDevices.size()-1]) ? "" : ",");
}
ov::CompiledModel compileModel = core.compile_model(model, "MULTI",
    ov::device::priorities(all_devices));

Beyond the trivial “CPU”, “GPU”, “HDDL” and so on, when multiple instances of a device are available the names are more qualified. For example, this is how two Intel® Movidius™ Myriad™ X sticks are listed with the hello_query_sample:

...
    Device: MYRIAD.1.2-ma2480
...
    Device: MYRIAD.1.4-ma2480

So the explicit configuration to use both would be “MULTI:MYRIAD.1.2-ma2480,MYRIAD.1.4-ma2480”. Accordingly, the code that loops over all available devices of “MYRIAD” type only is below:

ov::Core core;
std::vector<std::string> myriadDevices = core.get_property("MYRIAD", ov::available_devices);
std::string all_devices;
for (size_t i = 0; i < myriadDevices.size(); ++i) {
    all_devices += std::string("MYRIAD.")
                            + myriadDevices[i]
                            + std::string(i < (myriadDevices.size() -1) ? "," : "");
}
ov::CompiledModel compileModel = core.compile_model("sample.xml", "MULTI",
    ov::device::priorities(all_devices));

Configuring the Individual Devices and Creating the Multi-Device On Top

As discussed in the first section, you shall configure each individual device as usual and then just create the “MULTI” device on top:

// configure the HDDL device first
ov::Core core;
std::shared_ptr<ov::Model> model = core.read_model("sample.xml");
// compile the modle on the multi-device,
// while specifying the configuration (devices along with priorities
// and the configuration of devices):
ov::CompiledModel compileModel = core.compile_model(model, "MULTI",
    ov::device::priorities("HDDL", "GPU"),
    ov::device::properties("HDDL", hddl_config),
    ov::device::properties("GPU", gpu_config));

// query the optimal number of requests:
uint32_t nireq = compileModel.get_property(ov::optimal_number_of_infer_requests);

An alternative is to combine all the individual device settings into a single config file and load that, allowing the Multi-Device plugin to parse and apply settings to the right devices. See the code example in the next section.

Note that while the performance of accelerators combines really well with Multi-Device, the CPU+GPU execution poses some performance caveats, as these devices share the power, bandwidth and other resources. For example it is recommended to enable the GPU throttling hint (which save another CPU thread for the CPU inference). See the Using the Multi-Device with OpenVINO samples and benchmarking the performance section below.

Querying the Optimal Number of Inference Requests

You can use the configure devices to query the optimal number of requests. Similarly, when using the Multi-Device you don’t need to sum over included devices yourself, you can query property directly:

ov::Core core;
ov::CompiledModel compileModel = core.compile_model("sample.xml", "MULTI:HDDL,GPU");
// query the optimal number of requests
uint32_t nireq = compileModel.get_property(ov::optimal_number_of_infer_requests);

Using the Multi-Device with OpenVINO Samples and Benchmarking the Performance

Every OpenVINO sample that supports the -d (which stands for “device”) command-line option transparently accepts Multi-Device. The Benchmark Application is the best reference for the optimal usage of Multi-Device. As discussed earlier, you do not need to set up the number of requests, CPU streams or threads because the application provides optimal performance out of the box. Below is an example command to evaluate HDDL+GPU performance with that:

./benchmark_app d MULTI:HDDL,GPU m <model> -i <input> -niter 1000

The Multi-Device plugin supports FP16 IR files. The CPU plugin automatically upconverts it to FP32 and the other devices support it natively. Note that no demos are (yet) fully optimized for Multi-Device, by means of supporting the ov::optimal_number_of_infer_requests property, using the GPU streams/throttling, and so on.

Video: MULTI Plugin

Performance Considerations for the Multi-Device Execution

This section covers few recommendations for the multi-device execution (applicable for both Python and C++):

  • MULTI usually performs best when the fastest device is specified first in the list of the devices. This is particularly important when the request-level parallelism is not sufficient (e.g. the number of request in the flight is not enough to saturate all devices).

  • Just like with any throughput-oriented execution, it is highly recommended to query the optimal number of inference requests directly from the instance of the ov:compiled_model. Please refer to the code of the benchmark_app, that exists in both C++ and Python, for more details.

  • Notice that for example CPU+GPU execution performs better with certain knobs which you can find in the code of the same Benchmark App sample. One specific example is disabling GPU driver polling, which in turn requires multiple GPU streams to amortize slower communication of inference completion from the device to the host.

  • Multi-device logic always attempts to save on the (e.g. inputs) data copies between device-agnostic, user-facing inference requests and device-specific ‘worker’ requests that are being actually scheduled behind the scene. To facilitate the copy savings, it is recommended to run the requests in the order that they were created.

Introducing the Multi-Device Plugin (Python)

Python

The Multi-Device plugin automatically assigns inference requests to available computational devices to execute the requests in parallel. By contrast, the Heterogeneous plugin can run different layers on different devices but not in parallel. The potential gains with the Multi-Device plugin are:

  • Improved throughput from using multiple devices (compared to single-device execution)

  • More consistent performance, since the devices share the inference burden (if one device is too busy, another can take more of the load)

Note that with Multi-Device the application logic is left unchanged, so you don’t need to explicitly compile the model on every device, create and balance the inference requests and so on. From the application point of view, this is just another device that handles the actual machinery. The only thing that is required to leverage performance is to provide the multi-device (and hence the underlying devices) with enough inference requests to process. For example, if you were processing 4 cameras on the CPU (with 4 inference requests), it might be desirable to process more cameras (with more requests in flight) to keep CPU and GPU busy via Multi-Device.

The setup of Multi-Device can be described in three major steps:

  1. Configure each device (using the conventional configure devices method

  2. Compile the model on the Multi-Device plugin created on top of a (prioritized) list of the configured devices. This is the only change needed in the application.

  3. As with any other CompiledModel call (resulting from compile_model), you create as many requests as needed to saturate the devices.

These steps are covered below in detail.

Defining and Configuring the Multi-Device Plugin

Following the OpenVINO™ convention of labeling devices, the Multi-Device plugin uses the name “MULTI”. The only configuration option for the Multi-Device plugin is a prioritized list of devices to use:

Parameter name

Parameter values

Default

Description

“MULTI_DEVICE_PRIORITIES”

comma-separated device names with no spaces

N/A

Prioritized list of devices

You can set the configuration directly as a string, or use the metric key MULTI_DEVICE_PRIORITIES from the multi/multi_device_config.hpp file, which defines the same string.

The Three Ways to Specify Devices Targets for the MULTI plugin

  • Option 1 - Pass a Prioritized List as a Parameter in ie.load_network()

    core = Core()

    # Read a network in IR or ONNX format
    model = core.read_model(model_path)
    compiled_model = core.compile_model(model=model, device_name="MULTI:CPU,GPU")
  • Option 2 - Pass a List as a Parameter, and Dynamically Change Priorities during Execution Notice that the priorities of the devices can be changed in real time for the compiled model:

    core = Core()

    # Read a network in IR or ONNX format
    model = core.read_model(model_path)
    core.set_property(device_name="MULTI", properties={"MULTI_DEVICE_PRIORITIES":"HDDL,GPU"})
    # Change priorities
    core.set_property(device_name="MULTI", properties={"MULTI_DEVICE_PRIORITIES":"GPU,HDDL"})
    core.set_property(device_name="MULTI", properties={"MULTI_DEVICE_PRIORITIES":"GPU"})
    core.set_property(device_name="MULTI", properties={"MULTI_DEVICE_PRIORITIES":"HDDL,GPU"})
    core.set_property(device_name="MULTI", properties={"MULTI_DEVICE_PRIORITIES":"CPU,HDDL,GPU"})
  • Option 3 - Use Explicit Hints for Controlling Request Numbers Executed by Devices There is a way to specify the number of requests that Multi-Device will internally keep for each device. If the original app was running 4 cameras with 4 inference requests, it might be best to share these 4 requests between 2 devices used in the MULTI. The easiest way is to specify a number of requests for each device using parentheses: “MULTI:CPU(2),GPU(2)” and use the same 4 requests in the app. However, such an explicit configuration is not performance-portable and not recommended. The better way is to configure the individual devices and query the resulting number of requests to be used at the application level. See Configuring the Individual Devices and Creating the Multi-Device On Top.

Enumerating Available Devices

The OpenVINO Runtime API features a dedicated methods to enumerate devices and their capabilities. See the Hello Query Device Python Sample. This is example output from the sample (truncated to device names only):

./hello_query_device
Available devices:
    Device: CPU
...
    Device: GPU.0
...
    Device: GPU.1
...
    Device: HDDL

A simple programmatic way to enumerate the devices and use with the multi-device is as follows:

    all_devices = "MULTI:"
    core = Core()
    model = core.read_model(model_path)
    all_devices += ",".join(core.available_devices)
    compiled_model = core.compile_model(model=model, device_name=all_devices)

Beyond the trivial “CPU”, “GPU”, “HDDL” and so on, when multiple instances of a device are available the names are more qualified. For example, this is how two Intel® Movidius™ Myriad™ X sticks are listed with the hello_query_sample:

...
    Device: MYRIAD.1.2-ma2480
...
    Device: MYRIAD.1.4-ma2480

So the explicit configuration to use both would be “MULTI:MYRIAD.1.2-ma2480,MYRIAD.1.4-ma2480”. Accordingly, the code that loops over all available devices of “MYRIAD” type only is below:

    match_list = []
    all_devices = "MULTI:"
    dev_match_str = "MYRIAD"
    core = Core()
    model = core.read_model(model_path)
    for d in core.available_devices:
        if dev_match_str in d:
            match_list.append(d)
    all_devices += ",".join(match_list)
    compiled_model = core.compile_model(model=model, device_name=all_devices)

Configuring the Individual Devices and Creating the Multi-Device On Top

It is possible to configure each individual device as usual and then create the “MULTI” device on top:

    core = Core()
    cpu_config = {}
    gpu_config = {}
    model = core.read_model(model_path)
    core.set_property(device_name="CPU", properties=cpu_config)
    core.set_property(device_name="GPU", properties=gpu_config)
    compiled_model = core.compile_model(model=model, device_name="MULTI:GPU,CPU")
    # Query the optimal number of requests
    nireq = compiled_model.get_property("OPTIMAL_NUMBER_OF_INFER_REQUESTS")

An alternative is to combine all the individual device settings into a single config file and load that, allowing the Multi-Device plugin to parse and apply settings to the right devices. See the code example in the next section.

Note that while the performance of accelerators works well with Multi-Device, the CPU+GPU execution poses some performance caveats, as these devices share power, bandwidth and other resources. For example it is recommended to enable the GPU throttling hint (which saves another CPU thread for CPU inferencing). See the section below titled Using the Multi-Device with OpenVINO Samples and Benchmarking the Performance.

Using the Multi-Device with OpenVINO Samples and Benchmarking the Performance

Every OpenVINO sample that supports the -d (which stands for “device”) command-line option transparently accepts Multi-Device. The Benchmark application is the best reference for the optimal usage of Multi-Device. As discussed earlier, you do not need to set up the number of requests, CPU streams or threads because the application provides optimal performance out of the box. Below is an example command to evaluate CPU+GPU performance with the Benchmark application:

benchmark_app d MULTI:CPU,GPU m <model>

The Multi-Device plugin supports FP16 IR files. The CPU plugin automatically upconverts it to FP32 and the other devices support it natively. Note that no demos are (yet) fully optimized for Multi-Device, by means of supporting the OPTIMAL_NUMBER_OF_INFER_REQUESTS metric, using the GPU streams/throttling, and so on.

Video: MULTI Plugin

Note

This video is currently available only for C++, but many of the same concepts apply to Python.