14. dpdk-test-mldev Application

The dpdk-test-mldev tool is a Data Plane Development Kit (DPDK) application that allows testing various mldev use cases. This application has a generic framework to add new mldev based test cases to verify functionality and measure the performance of inference execution on DPDK ML devices.

14.1. Application and Options

The application has a number of command line options:

dpdk-test-mldev [EAL Options] -- [application options]

14.1.1. EAL Options

The following are the EAL command-line options that can be used with the dpdk-test-mldev application. See the DPDK Getting Started Guides for more information on these options.

-c <COREMASK> or -l <CORELIST>
Set the hexadecimal bitmask of the cores to run on. The corelist is a list of cores to use.
-a <PCI_ID>
Attach a PCI based ML device. Specific to drivers using a PCI based ML device.
--vdev <driver>
Add a virtual mldev device. Specific to drivers using a ML virtual device.

14.1.2. Application Options

The following are the command-line options supported by the test application.

--test <name>

Name of the test to execute. ML tests are divided into three groups: Device, Model and Inference tests. Test name should be one of the following supported tests.

ML Device Tests

device_ops

ML Model Tests

model_ops

ML Inference Tests

inference_ordered
inference_interleave
--dev_id <n>
Set the device ID of the ML device to be used for the test. Default value is 0.
--socket_id <n>
Set the socket ID of the application resources. Default value is SOCKET_ID_ANY.
--models <model_list>
Set the list of model files to be used for the tests. Application expects the model_list in comma separated form (i.e. --models model_A.bin,model_B.bin). Maximum number of models supported by the test is 8.
--filelist <file_list>

Set the list of model, input, output and reference files to be used for the tests. Application expects the file_list to be in comma separated form (i.e. --filelist <model,input,output>[,reference]).

Multiple filelist entries can be specified when running the tests with multiple models. Both quantized and dequantized outputs are written to the disk. Dequantized output file would have the name specified by the user through --filelist option. A suffix .q is appended to quantized output filename. Maximum number of filelist entries supported by the test is 8.

--repetitions <n>
Set the number of inference repetitions to be executed in the test per each model. Default value is 1.
--burst_size <n>
Set the burst size to be used when enqueuing / dequeuing inferences. Default value is 1.
--queue_pairs <n>
Set the number of queue-pairs to be used for inference enqueue and dequeue operations. Default value is 1.
--queue_size <n>
Set the size of queue-pair to be created for inference enqueue / dequeue operations. Queue size would translate into rte_ml_dev_qp_conf::nb_desc field during queue-pair creation. Default value is 1.
--batches <n>
Set the number batches in the input file provided for inference run. When not specified, the test would assume the number of batches is the batch size of the model.
--tolerance <n>
Set the tolerance value in percentage to be used for output validation. Default value is 0.
--stats
Enable reporting device extended stats.
--debug
Enable the tests to run in debug mode.
--help
Print help message.

14.2. ML Device Tests

ML device tests are functional tests to validate ML device API. Device tests validate the ML device handling configure, close, start and stop APIs.

14.2.1. Application Options

Supported command line options for the device_ops test are following:

--debug
--test
--dev_id
--socket_id
--queue_pairs
--queue_size

14.2.2. DEVICE_OPS Test

Device ops test validates the device configuration and reconfiguration support. The test configures ML device based on the options --queue_pairs and --queue_size specified by the user, and later reconfigures the ML device with the number of queue pairs and queue size based on the maximum specified through the device info.

14.2.2.1. Example

Command to run device_ops test:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=device_ops

Command to run device_ops test with user options:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=device_ops --queue_pairs <M> --queue_size <N>

14.3. ML Model Tests

Model tests are functional tests to validate ML model API. Model tests validate the functioning of load, start, stop and unload ML models.

14.3.1. Application Options

Supported command line options for the model_ops test are following:

--debug
--test
--dev_id
--socket_id
--models

List of model files to be used for the model_ops test can be specified through the option --models <model_list> as a comma separated list. Maximum number of models supported in the test is 8.

Note

  • The --models <model_list> is a mandatory option for running this test.
  • Options not supported by the test are ignored if specified.

14.3.2. MODEL_OPS Test

The test is a collection of multiple sub-tests, each with a different order of slow-path operations when handling with N number of models.

Sub-test A: executes the sequence of load / start / stop / unload for a model in order, followed by next model.

../_images/mldev_model_ops_subtest_a.svg

Fig. 14.1 Execution sequence of model_ops subtest A.

Sub-test B: executes load for all models, followed by a start for all models. Upon successful start of all models, stop is invoked for all models followed by unload.

../_images/mldev_model_ops_subtest_b.svg

Fig. 14.2 Execution sequence of model_ops subtest B.

Sub-test C: loads all models, followed by a start and stop of all models in order. Upon completion of stop, unload is invoked for all models.

../_images/mldev_model_ops_subtest_c.svg

Fig. 14.3 Execution sequence of model_ops subtest C.

Sub-test D: executes load and start for all models available. Upon successful start of all models, stop is executed for the models.

../_images/mldev_model_ops_subtest_d.svg

Fig. 14.4 Execution sequence of model_ops subtest D.

14.3.2.1. Example

Command to run model_ops test:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=model_ops --models model_1.bin,model_2.bin,model_3.bin, model_4.bin

14.4. ML Inference Tests

Inference tests are a set of tests to validate end-to-end inference execution on ML device. These tests executes the full sequence of operations required to run inferences with one or multiple models.

14.4.1. Application Options

Supported command line options for inference tests are following:

--debug
--test
--dev_id
--socket_id
--filelist
--repetitions
--burst_size
--queue_pairs
--queue_size
--batches
--tolerance
--stats

List of files to be used for the inference tests can be specified through the option --filelist <file_list> as a comma separated list. A filelist entry would be of the format --filelist <model_file,input_file,output_file>[,reference_file] and is used to specify the list of files required to test with a single model. Multiple filelist entries are supported by the test, one entry per model. Maximum number of file entries supported by the test is 8.

When --burst_size <num> option is specified for the test, enqueue and dequeue burst would try to enqueue or dequeue num number of inferences per each call respectively.

In the inference test, a pair of lcores are mapped to each queue pair. Minimum number of lcores required for the tests is equal to (queue_pairs * 2 + 1).

Output validation of inference would be enabled only when a reference file is specified through the --filelist option. Application would additionally consider the tolerance value provided through --tolerance option during validation. When the tolerance values is 0, CRC32 hash of inference output and reference output are compared. When the tolerance is non-zero, element wise comparison of output is performed. Validation is considered as successful only when all the elements of the output tensor are with in the tolerance range specified.

Enabling --stats would print the extended stats supported by the driver.

Note

  • The --filelist <file_list> is a mandatory option for running inference tests.
  • Options not supported by the tests are ignored if specified.
  • Element wise comparison is not supported when the output dtype is either fp8, fp16 or bfloat16. This is applicable only when the tolerance is greater than zero and for pre-quantized models only.

14.4.2. INFERENCE_ORDERED Test

This is a functional test for validating the end-to-end inference execution on ML device. This test configures ML device and queue pairs as per the queue-pair related options (queue_pairs and queue_size) specified by the user. Upon successful configuration of the device and queue pairs, the first model specified through the filelist is loaded to the device and inferences are enqueued by a pool of worker threads to the ML device. Total number of inferences enqueued for the model are equal to the repetitions specified. A dedicated pool of worker threads would dequeue the inferences from the device. The model is unloaded upon completion of all inferences for the model. The test would continue loading and executing inference requests for all models specified through filelist option in an ordered manner.

../_images/mldev_inference_ordered.svg

Fig. 14.5 Execution of inference_ordered on single model.

14.4.2.1. Example

Example command to run inference_ordered test:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_ordered --filelist model.bin,input.bin,output.bin

Example command to run inference_ordered test with a specific burst size:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_ordered --filelist model.bin,input.bin,output.bin \
     --burst_size 12

Example command to run inference_ordered test with multiple queue-pairs and queue size:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_ordered --filelist model.bin,input.bin,output.bin \
     --queue_pairs 4 --queue_size 16

Example command to run inference_ordered with output validation using tolerance of 1%:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_ordered --filelist model.bin,input.bin,output.bin,reference.bin \
     --tolerance 1.0

14.4.3. INFERENCE_INTERLEAVE Test

This is a stress test for validating the end-to-end inference execution on ML device. The test configures the ML device and queue pairs as per the queue-pair related options (queue_pairs and queue_size) specified by the user. Upon successful configuration of the device and queue pairs, all models specified through the filelist are loaded to the device. Inferences for multiple models are enqueued by a pool of worker threads in parallel. Inference execution by the device is interleaved between multiple models. Total number of inferences enqueued for a model are equal to the repetitions specified. An additional pool of threads would dequeue the inferences from the device. Models would be unloaded upon completion of inferences for all models loaded.

../_images/mldev_inference_interleave.svg

Fig. 14.6 Execution of inference_interleave on single model.

14.4.3.1. Example

Example command to run inference_interleave test:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_interleave --filelist model.bin,input.bin,output.bin

Example command to run inference_interleave test with multiple models:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_interleave --filelist model_A.bin,input_A.bin,output_A.bin \
     --filelist model_B.bin,input_B.bin,output_B.bin

Example command to run inference_interleave test with a specific burst size, multiple queue-pairs and queue size:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_interleave --filelist model.bin,input.bin,output.bin \
     --queue_pairs 8 --queue_size 12 --burst_size 16

Example command to run inference_interleave test with multiple models and output validation using tolerance of 2.0%:

sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
     --test=inference_interleave \
     --filelist model_A.bin,input_A.bin,output_A.bin,reference_A.bin \
     --filelist model_B.bin,input_B.bin,output_B.bin,reference_B.bin \
     --tolerance 2.0

14.5. Debug mode

ML tests can be executed in debug mode by enabling the option --debug. Execution of tests in debug mode would enable additional prints.

When a validation failure is observed, output from that buffer is written to the disk, with the filenames having similar convention when the test has passed. Additionally index of the buffer would be appended to the filenames.