Triton Server Trace

Triton includes that capability to generate a detailed trace for individual inference requests. Tracing is enable by command-line arguments when running the tritonserver executable.

--trace-config command line option in Triton can be used to specify global and trace mode specific config setting. The format of this flag is --trace-config <mode>,<setting>=<value>, where <mode> is either triton or opentelemetry. By default, the trace mode is set to triton, and the server will use Triton's trace APIs. For opentelemetry mode, the server will use the OpenTelemetry's APIs to generate, collect and export traces for individual inference requests.

To specify global trace settings (level, rate, count, or mode), the format is --trace-config <setting>=<value>.

An example usage, which invokes Triton's trace APIs:

$ tritonserver \
    --trace-config triton,file=/tmp/trace.json \
    --trace-config triton,log-frequency=50 \
    --trace-config rate=100 \
    --trace-config level=TIMESTAMPS \
    --trace-config count=100 ...

Trace Settings

Global Settings

The following table shows available global trace settings to pass to --trace-config

Setting	Default Value	Description
`rate`	1000	Specifies the sampling rate. The same as deprecated `--trace-rate`. For example, a value of 1000 specifies that every 1000-th inference request will be traced.
`level`	OFF	Indicates the level of trace detail that should be collected and may be specified multiple times to trace multiple information. The same as deprecated `--trace-level`. Choices are `TIMESTAMPS` and `TENSORS`. Note that `opentelemetry` mode does not currently support `TENSORS` level.
`count`	-1	Specifies the remaining number of traces to be collected. The default value of -1 specifies to never stop collecting traces. With a value of 100, Triton will stop tracing requests after 100 traces are collected. The same as deprecated `--trace-count`.
`mode`	triton	Specifies which trace APIs to use for collecting traces. The choices are `triton` or `opentelemetry`.

Triton Trace APIs Settings

The following table shows available Triton trace APIs settings for --trace-config triton,<setting>=<value>.

Setting	Default Value	Description
`file`	empty string	Indicates where the trace output should be written. The same as deprecated `--trace-file`.
`log-frequency`	0	Specifies the rate that the traces are written to file. For example, a value of 50 specifies that Triton will log to file for every 50 traces collected. The same as deprecated `--trace-log-frequency`.

In addition to the trace configuration settings in the command line, you can modify the trace configuration using the trace protocol. This option is currently not supported, when trace mode is set to opentelemetry.

Note: the following flags are deprecated:

The --trace-file option indicates where the trace output should be written. The --trace-rate option specifies the sampling rate. In this example every 100-th inference request will be traced. The --trace-level option indicates the level of trace detail that should be collected. --trace-level option may be specified multiple times to trace multiple information. The --trace-log-frequency option specifies the rate that the traces are written to file. In this example Triton will log to file for every 50 traces collected. The --trace-count option specifies the remaining number of traces to be collected. In this example Triton will stop tracing more requests after 100 traces are collected. Use the --help option to get more information.

Supported Trace Level Option

TIMESTAMPS: Tracing execution timestamps of each request.
TENSORS: Tracing input and output tensors during the execution.

JSON Trace Output

The trace output is a JSON file with the following schema.

[
  {
    "model_name": $string,
    "model_version": $number,
    "id": $number,
    "request_id": $string,
    "parent_id": $number
  },
  {
    "id": $number,
    "timestamps": [
      { "name" : $string, "ns" : $number }
    ]
  },
  {
    "id": $number
    "activity": $string,
    "tensor":{
      "name": $string,
      "data": $string,
      "shape": $string,
      "dtype": $string
    }
  },
  ...
]

Each trace is assigned a "id", which indicates the model name and version of the inference request. If the trace is from a model run as part of an ensemble, the "parent_id" will indicate the "id" of the containing ensemble. For example:

[
  {
    "id": 1,
    "model_name": "simple",
    "model_version": 1
  },
  ...
]

Each TIMESTAMPS trace will have one or more "timestamps" with each timestamp having a name and the timestamp in nanoseconds ("ns"). For example:

[
  {"id": 1, "timestamps": [{ "name": "HTTP_RECV_START", "ns": 2356425054587444 }] },
  {"id": 1, "timestamps": [{ "name": "HTTP_RECV_END", "ns": 2356425054632308 }] },
  {"id": 1, "timestamps": [{ "name": "REQUEST_START", "ns": 2356425054785863 }] },
  {"id": 1, "timestamps": [{ "name": "QUEUE_START", "ns": 2356425054791517 }] },
  {"id": 1, "timestamps": [{ "name": "INFER_RESPONSE_COMPLETE", "ns": 2356425057587919 }] },
  {"id": 1, "timestamps": [{ "name": "COMPUTE_START", "ns": 2356425054887198 }] },
  {"id": 1, "timestamps": [{ "name": "COMPUTE_INPUT_END", "ns": 2356425057152908 }] },
  {"id": 1, "timestamps": [{ "name": "COMPUTE_OUTPUT_START", "ns": 2356425057497763 }] },
  {"id": 1, "timestamps": [{ "name": "COMPUTE_END", "ns": 2356425057540989 }] },
  {"id": 1, "timestamps": [{ "name": "REQUEST_END", "ns": 2356425057643164 }] },
  {"id": 1, "timestamps": [{ "name": "HTTP_SEND_START", "ns": 2356425057681578 }] },
  {"id": 1, "timestamps": [{ "name": "HTTP_SEND_END", "ns": 2356425057712991 }] }
]

Each TENSORS trace will contain an "activity" and a "tensor". "activity" indicates the type of tensor, including "TENSOR_QUEUE_INPUT" and "TENSOR_BACKEND_OUTPUT" by now. "tensor" has the detail of tensor, including its "name", "data" and "dtype". For example:

[
  {
    "id": 1,
    "activity": "TENSOR_QUEUE_INPUT",
    "tensor":{
      "name": "input",
      "data": "0.1,0.1,0.1,...",
      "shape": "1,16",
      "dtype": "FP32"
    }
  }
]

Trace Summary Tool

An example trace summary tool can be used to summarize a set of traces collected from Triton. Basic usage is:

$ trace_summary.py <trace file>

This produces a summary report for all traces in the file. HTTP and GRPC inference requests are reported separately.

File: trace.json
Summary for simple (-1): trace count = 1
HTTP infer request (avg): 403.578us
	Receive (avg): 20.555us
	Send (avg): 4.52us
	Overhead (avg): 24.592us
	Handler (avg): 353.911us
  		Overhead (avg): 23.675us
  		Queue (avg): 18.019us
  		Compute (avg): 312.217us
  			Input (avg): 24.151us
  			Infer (avg): 244.186us
  			Output (avg): 43.88us
Summary for simple (-1): trace count = 1
GRPC infer request (avg): 383.601us
	Send (avg): 62.816us
	Handler (avg): 392.924us
  		Overhead (avg): 51.968us
  		Queue (avg): 21.45us
  		Compute (avg): 319.506us
  			Input (avg): 27.76us
  			Infer (avg): 227.844us
  			Output (avg): 63.902us

Note: The "Receive (avg)" metric is not included in the gRPC summary as gRPC library does not provide any non-intrusive hooks to detect time spent in reading a message from the wire. Tracing an HTTP request will provide an accurate measurement of time spent reading a request from the network.

Use the -t option to get a summary for each trace in the file. This summary shows the time, in microseconds, between different points in the processing of an inference request. For example, the below output shows that it took 15us from the start of handling the request until the request was enqueued in the scheduling queue.

$ trace_summary.py -t <trace file>
...
simple (-1):
  	request handler start
  		15us
  	queue start
  		20us
  	compute start
  		266us
  	compute end
  		4us
  	request handler end
  		19us
  	grpc send start
  		77us
  	grpc send end
...

The script can also show the data flow of the first request if there are TENSORS traces in the file. If the TENSORS traces are from an ensemble, the data flow will be shown with the dependency of each model.

...
Data Flow:
	==========================================================
	Name:   ensemble
	Version:1
	QUEUE_INPUT:
		input: [[0.705676  0.830855  0.833153]]
	BACKEND_OUTPUT:
		output: [[1. 2. 7. 0. 4. 7. 9. 3. 4. 9.]]
	==========================================================
		==================================================
		Name:   test_trt1
		Version:1
		QUEUE_INPUT:
			input: [[0.705676  0.830855  0.833153]]
		BACKEND_OUTPUT:
			output1: [[1. 1. ...]]
		==================================================
		==================================================
		Name:   test_trt2
		Version:1
		QUEUE_INPUT:
			input: [[0.705676  0.830855  0.833153]]
		BACKEND_OUTPUT:
			output2: [[2. 2. ...]]
		==================================================
		==================================================
		Name:   test_py
		Version:1
		QUEUE_INPUT:
			output1: [[1. 1. ...]]
		QUEUE_INPUT:
			output2: [[2. 2. ...]]
		BACKEND_OUTPUT:
			output: [[1. 2. 7. 0. 4. 7. 9. 3. 4. 9.]]
		==================================================
...

The meaning of the trace timestamps is:

HTTP Request Receive: Collected only for inference requests that use the HTTP protocol. The time required to read the inference request from the network.
Send: The time required to send the inference response.
Overhead: Additional time required in the HTTP endpoint to process the inference request and response.
Handler: The total time spent handling the inference request, not including the HTTP and GRPC request/response handling.
- Queue: The time the inference request spent in the scheduling queue.
- Compute: The time the inference request spent executing the actual inference. This time includes the time spent copying input and output tensors. If --trace-level=TIMESTAMPS then a breakdown of the compute time will be provided as follows:
  - Input: The time to copy input tensor data as required by the inference framework / backend. This includes the time to copy input tensor data to the GPU.
  - Infer: The time spent executing the model to perform the inference.
  - Output: The time to copy output tensor data as required by the inference framework / backend. This includes the time to copy output tensor data from the GPU.
- Overhead: Additional time required for request handling not covered by Queue or Compute times.
Data Flow: The data flow of the first request. It contains the input and output tensors of each part of execution.
- Name: The name of model.
- Version: The version of model.
- QUEUE_INPUT: The tensor entering the queue of a backend to wait for scheduling.
- BACKEND_OUTPUT: The tensor in the response of a backend.

Tracing for BLS models

Triton does not collect traces for child models invoked from BLS models by default.

To include child models into collected traces, user needs to provide the trace argument (as shown in the example below), when constructing an InferenceRequest object. This helps Triton associate the child model with the parent model's trace (request.trace()).

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
  ...
    def execute(self, requests):
      ...
      for request in requests:
        ...
        inference_request = pb_utils.InferenceRequest(
            model_name='model_name',
            requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
            inputs=[<pb_utils.Tensor object>], trace = request.trace())

OpenTelemetry trace support

Triton provides an option to generate and export traces using OpenTelemetry APIs and SDKs.

To specify OpenTelemetry mode for tracing, specify the --trace-config flag as follows:

$ tritonserver --trace-config mode=opentelemetry \
    --trace-config opentelemetry,url=<endpoint> ...

Triton's OpenTelemetry trace mode uses Batch Span Processor, which batches ended spans and sends them in bulk. Batching helps with data compression and reduces the number of outgoing connections required to transmit the data. This processor supports both size and time based batching. Size-based batching is controlled by 2 parameters: bsp_max_export_batch_size and bsp_max_queue_size, while time-based batching is controlled by bsp_schedule_delay. Collected spans will be exported when the batch size reaches bsp_max_export_batch_size, or delay since last export reaches bsp_schedule_delay, whatever comes first. Additionally, user should make sure that bsp_max_export_batch_size is always less than bsp_max_queue_size, otherwise the excessive spans will be dropped and trace data will be lost.

Default parameters for the Batch Span Processor are provided in OpenTelemetry trace APIs settings. As a general recommendation, make sure that bsp_max_queue_size is large enough to hold all collected spans, and bsp_schedule_delay does not cause frequent exports, which will affect Triton Server's latency. A minimal Triton trace consists of 3 spans: top level span, model span, and compute span.

Top level span: The top-level span collects timestamps for when request was received by Triton, and when the response was sent. Any Triton trace contains only 1 top level span.
Model span: Model spans collect information, when request for this model was started, when it was placed in a queue, and when it was ended. A minimal Triton trace contains 1 model span.
Compute span: Compute spans record compute timestamps. A minimal Triton trace contains 1 compute span.

The total amount of spans depends on the complexity of your model. A general rule is any base model - a single model that performs computations - produces 1 model span and one compute span. For ensembles, every composing model produces model and compute spans in addition to one model span for the ensemble. BLS models produce the same number of model and compute spans as the total amount of models involved in the BLS request, including the main BLS model.

Differences in trace contents from Triton's trace output

OpenTelemetry APIs produce spans that collect the same timestamps as Triton's Trace APIs. Each span also includes model_name, model_version, request_id, and parent_id as an attribute.

The span collects TIMESTAMPS that consist of a name and a timestamp in nanoseconds, which is similar to Triton Trace APIs. However, OpenTelemetry relies on the system's clock for event timestamps, which is based on the system's real-time clock. On the other hand, Triton Trace APIs report timestamps using steady clock, which is a monotonic clock that ensures time always movess forward. This clock is not related to wall clock time and, for example, can measure time since last reboot.

OpenTelemetry trace APIs settings

The following table shows available OpenTelemetry trace APIs settings for --trace-config opentelemetry,<setting>=<value>.

Setting	Default Value	Description
`url`	`http://localhost:4318/v1/traces`	`host:port` to which the receiver is going to receive trace data.
`resource`	`service.name=triton-inference-server`	Key-value pairs to be used as resource attributes. Should be specified following the provided template: `--trace-config opentelemetry,resource=<key>=<value>` For example: `--trace-config opentelemetry,resource=service.name=triton` `--trace-config opentelemetry,resource=service.version=1` Alternatively, key-value attributes can be specified through OTEL_RESOURCE_ATTRIBUTES environment variable.
Batch Span Processor
`bsp_max_queue_size`	2048	Maximum queue size. This setting can also be specified through OTEL_BSP_MAX_QUEUE_SIZE environment variable.
`bsp_schedule_delay`	5000	Delay interval (in milliseconds) between two consecutive exports. This setting can also be specified through OTEL_BSP_SCHEDULE_DELAY environment variable.
`bsp_max_export_batch_size`	512	Maximum batch size. Must be less than or equal to `bsp_max_queue_size`. This setting can also be specified through OTEL_BSP_MAX_EXPORT_BATCH_SIZE environment variable.

OpenTelemetry Context Propagation

Triton supports context propagation in OpenTelemetry mode starting in version 24.01. Note, that every request with propagated OpenTelemetry context will be traced, regardless of rate and count trace settings. If a user wishes to trace only those requests, for which OpenTelemetry context was injected on the client side, please start Triton with --trace-config rate=0:

$ tritonserver \
    --trace-config rate=0 \
    --trace-config level=TIMESTAMPS \
    --trace-config count=-1 \
    --trace-config mode=opentelemetry

Please, be aware that this option is subject to change in future releases.

How to inject OpenTelemetry context on the client side

For C++ clients, please refer to gRPC and HTTP examples.

For python clients, please make sure to install OpenTelemetry Python. You can then use the opentelemetry.propagate.inject method to prepare headers to pass with the request, as shown here. Then, you can specify headers in the infer method. For references, please look at our tests, e.g. http context propagation test.

Custom Backend Tracing

In the case when a custom activity needs to be traced in the backend, please use TRITONSERVER_InferenceTraceReportActivity API. For examples, please refer to the identity backend.

In openTelemetry trace mode, if one wishes to start a new span, make sure that the name of your custom activity ends with _START. To end the new span, make sure that corresponding activity ends with _END. For example, in the identity backend, we start a CUSTOM_ACTIVITY span, by reporting CUSTOM_ACTIVITY_START event; and we close this span by reporting CUSTOM_ACTIVITY_END event.

Please note, that it is user's responsibility to make sure that all custom started spans are properly ended.

Limitations

OpenTelemetry trace mode is not supported on Windows systems.
Triton supports only OTLP/HTTP Exporter and allows specification of only url for this exporter through --trace-config. Other options and corresponding default values can be found here.
Triton does not support configuration of the opentelemetry trace settings during a Triton run and opentelemetry specific settings are not available for the retrieval through Triton's trace extension.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!