Explore analytics with SambaTune

After you’ve installed and run SambaTune, you can look at information about the model run, and then drill down into analytics.

Prerequisites

Ensure that SambaTune is installed in your environment. See Install SambaTune
Run sambatune, passing in your custom YAML file for the model you want to tune. See Run SambaTune.
Run sambatune_ui and note down the URL, username and password that you see at the end of the run. A random password is generated for each run. See Run the SambaTune GUI.

From your web browser, connect to the SambaTune client. You use the information that you saw when you ran sambatune_ui.
In the Dashboard, either use the selected directory or specify a different directory.
Select one of the subdirectories by clicking the checkbox and click View.
Click Analytics on the left.
With Host selected, explore performance on the host.
1. Examine the Diagnoses and Runtime statistics. See Performance Insight Report for details.
2. With latency selected, you get a summary of the information in the Host-Device Breakdown.
3. Change the sort order as needed.
4. Explore the Pyinstrument data.

Click RDU in the graph to learn more about performance on RDU. The different tabs allow detailed insight into different aspects of performance.

Tab	Information
Summary	Click the check boxes to focus on what you’re interested in.
Section Report	See Section Report documentation
Collated Report	See PMU Bandwidth Report documentation
Paths Report	Paths Report documentation

Tab

Information

Summary

Click the check boxes to focus on what you’re interested in.

Section Report

Collated Report

Paths Report

The Training/Inference Statistics tab is currently in Beta. Expect updates in a future release.

One option is to look at the workflow example, which includes screenshots.
You can compare analytics for two comparable model runs. See Compare performance metrics.