Introduction

The plugin performance statistics tool can quickly test and analyze the inference performance of the MindX SDK. (The time unit of all performance statistics is microsecond.) For details about performance metrics, see Table 1:

Figure 1 MindX SDK inference process

**Table 1** Performance metrics
Performance Metric	Description
End-to-end duration	The timing starts when the srcpad of the appsrc input plugin outputs data (t1) and ends when the sinkpad of the appsink output plugin receives data (t4). E2E time = End time (t4) – Start time (t1)
Single plugin duration	Collects the duration of a single plugin. Only SDK-developed plugins (plugins whose names start with mxpi_) are supported, while the GStreamer plugins are excluded. For example, the time consumed by the mxpi_modelinfer plugin starts from the sinkpad (at t2) and ends at the srcpad (at t3). Duration of a single plugin = End time (t3) – Start time (t2)
Model inference duration	Model inference is involved only in the model plugin. Therefore, the model inference duration is collected in this plugin.
Post-processing duration	Post-processing exists only in the model inference plugin of the old framework. Therefore, the postprocessing duration is collected in the plugin of the old framework.
Throughput	Collects statistics each time the appsink output plugin receives data from the sinkpad.
Queue length	Obtains the current queue length at an interval specified by ps_queue_size_interval_time (see Configuration Item Description). When the value of ps_queue_size_times (see Configuration Item Description) is obtained, the queue length is measured.

Parent topic: Plugin Performance Statistics Tool