Performance visualization

CARET serves Python APIs to visualize trace data for users to analyze applications' performance. This section will show basic flow to visualize data with Jupyter notebook.

How to locate trace data on Jupyter notebook

1. Launching jupyter-lab Launch jupyter-lab at first for learning how to visualize recorded data.

   cd ~/ros2_ws/evaluate
   
   source ~/ros2_caret_ws/install/setup.bash
   jupyter-lab
   

   CARET uses bokeh for some visualization APIs. Execute the following code for loading bokeh

   from bokeh.plotting import output_notebook
   output_notebook()
   

2. Loading trace data onto Jupyter notebook

   Locate trace data on Jupyter notebook as well as the architecture file.

   from caret_analyze import Architecture, Application, Lttng
   
   # load the architecture file which is created in the previous page
   arch = Architecture('yaml', './architecture.yaml')
   
   # load recorded data by CARET
   lttng = Lttng('./e2e_sample')
   
   # map the application architecture to recorded data
   app = Application(arch, lttng)
   

   After execution of the code, users will often refer to the app object defined as Application class. The app objects provides users latency of callbacks, communication, and paths. Application class is similar to Architecture class which describes structure of the application and their interfaces are similar. In addition, Application class has interfaces to get latency.

Basic API to get latency

This section will explain APIs to get callback latency. The following code is one of examples to get callback latency.

# Get a callback instance, which has latency information, from app
callback = app.get_callback('/timer_driven_node/callback_0')

# Get time-series variation of latency
t, latency_ns = callback.to_timeseries()

# Get histogram of latency
bins, latency_ns = callback.to_histogram()

Though this example shows callback latency, CARET serves API to get communication latency. callback.to_dataframe() will provides pandas.DataFrame based object including raw timestamps which is obtained from tracepoints.

Todo

Sorry for not providing CARET's API list, but we'll provide it in the near future.

Visualizing latency of node chain with message flow

CARET provides some APIs to visualize measured data. With message flow, which is one of effective visualizations, users can understand what happens and where is bottleneck in target applications. Execute the following code for visualization with message flow.

from caret_analyze.plot import Plot

target_path = app.get_path('target')
plot = Plot.create_message_flow_plot(target_path)
plot.show()

The following figure will appear if show method is successful.

The horizontal axis shows time line. On the other hand, the vertical axis shows elements of the node chain (target), which include callback functions and topic messages from input to output. Each colorful line shows how a certain input message is propagated to callback functions and topic communications. Gray rectangles show execution timing of callback functions.

If your mouse pointer is put over one of gray rectangles, latency of callback functions will be shown. You can find latency of a target node chain as well if you put the mouse pointer onto one of colorful lines.

CARET serves other APIs for visualization. Refer to visualization for more details.