pipeline_latency_monitor

Purpose

This package provides a node for calculating the latency introduced by a sequence of Autoware processing steps. For example, the pipeline from receiving sensor data up to the generation of the control command is made of several steps (e.g., perception, prediction, planning, control). This node calculates the total end-to-end latency by subscribing to messages that report the processing time of each step, combining them based on their timestamps, and publishing the total latency.

Inner-workings / Algorithms

The node is configured with a sequence of processing steps. For each step, it subscribes to a topic that publishes its execution latency.

To calculate the total latency, the node works backward from the last step in the defined sequence. It takes the most recent latency measurement from the final step. Then, for each preceding step, it finds a corresponding latency measurement from its history whose timestamp is consistent with the subsequent step. Specifically, it ensures that the end time of a given step occurs before the start time of the next step in the chain.

The total latency is the sum of these chronologically-consistent individual latencies. The node also allows for adding fixed offset values to the final sum.

The calculated total latency is published continuously and also used to report diagnostic information. A WARN status is published if the latency exceeds a configurable threshold.

Inputs / Outputs

Input

Subscribes to topics containing latency information for each processing step. The topic names and types are configured via parameters. Only 2 message types are currently supported.

Message Type	Timestamp field	Latency value field
autoware_internal_debug_msgs/msg/Float64Stamped	stamp	data
autoware_planning_validator/msg/PlanningValidatorStatus	stamp	latency

Output

Name	Type	Description
~/output/total_latency_ms	autoware_internal_debug_msgs/msg/Float64Stamped	The calculated total pipeline latency in milliseconds.
/diagnostics	diagnostic_msgs/DiagnosticArray	Publishes the diagnostic status. Reports OK if the latency is within the threshold, and WARN if it is exceeded.
~/debug/<step_name>_latency_ms	autoware_internal_debug_msgs/msg/Float64Stamped	For each processing step, publishes the latest latency value received from its input topic.
~/debug/pipeline_total_latency_ms	autoware_internal_debug_msgs/msg/Float64Stamped	A debug topic that also publishes the calculated total latency.

Parameters

Name	Type	Default Value	Description
update_rate	double	10	The rate [Hz] at which the total latency is calculated and published.
latency_threshold_ms	double	1000	The latency threshold in milliseconds. If the total latency exceeds this value, a WARN diagnostic is reported.
window_size	int	10	The number of historical latency messages to store for each processing step.
processing_steps.sequence	string[]	[]	An array of names defining the ordered sequence of processing steps to measure.
processing_steps.<step_name>.topic	string	-	The input topic that provides the latency for this step.
processing_steps.<step_name>.topic_type	string	-	The message type of the input topic. Supported types: autoware_internal_debug_msgs/msg/Float64Stamped, autoware_planning_validator/msg/PlanningValidatorStatus.
processing_steps.<step_name>.timestamp_meaning	string	"end"	Defines if the timestamp in the message represents the start or end of the processing interval for that step.
processing_steps.<step_name>.latency_multiplier	double	1.0	A multiplier to convert the received latency value into milliseconds (e.g., use 1000.0 if the input is in seconds).
latency_offsets_ms	double[]	[]	A list of fixed latency values (in milliseconds) to be added to the calculated total latency. Useful for accounting for steps that do not publish their own latency.

Assumptions / Known limits

• The accuracy of the latency calculation depends on the timestamps of all input messages being synchronized to a common clock.
• The algorithm for combining latencies assumes a sequential, non-parallel processing pipeline.
• The node must be configured with the correct sequence of processing steps and their corresponding topics for the calculation to be meaningful.