Starting VTune

$ cd /opt/intel/oneapi/vtune/latest/bin64
$ ./vtune-gui

Create a Project

Once the GUI has started, click on the New Project… button

Select a System to Run the Analysis

Click on Local Host to run the analysis on the local machine (as opposed to a remote machine).

Select an Analysis Target

• Select the analysis target. This can be an application or the entire system.
• The application can be a binary or a script.
• Your app may require parameters. Configure them under application parameters.

Performance Snapshot

This gives you a quick overview of potential performance issues.

• Open the menu to select analysis types.

Configure Binary and Source Directories

• At the bottom of the GUI, click on the folder icon to select the binary and source directories for post-processing if they do not correspond to debug information.

Generate a Command Line

• Once you have selected your options you can click on the console icon next to the folder icon to generate a command to run your analysis. This saves you from remembering all of the command line options.

Start the App and Start to Profile the App

• Click on the start button near the bottom of the GUI to start the application and begin the data collection process.

Monitor the Elapsed Time

The elapsed time, near the top right of the GUI is a great indicator of tuning efforts. Track this before and after making changes.

Choose your Next Analysis Type

Choose from many analysis types for your next analysis e.g.

• Algorithm
  • Hotspots: Algorithm efficiency
  • Anomaly Detection
  • Memory Consumption
• Parallelism
  • Threading
  • HPC Performance Characterization
• Accelerators
  • GPU Offload
  • GPU Compute/Media Hotspots
  • CPU/FPGA Interaction
• Microarchitecture
  • Microarchitecture Exploration
  • Memory Access
• I/O
  • Input and Output
• Platform Analysis
  • System Overview
  • GPU Rendering

Optimizations

• Try to optimize the top hotspot functions
• Click on a function to explore its stack
• Use the CPU Utilization Histogram to analyze the duration that a specific number of CPUs ran simultaneously.
  • Adjust sliders to set thresholds
• The top right corner is your performance navigator. Use it to plan next steps.

Calling Sequence

• Use the Bottom-up tab to explore each functions calling sequence: from a child up to a parent function.
• Expand a hotspot function to identify a call path with the highest CPU time.
• Group data with available hierarhies e.g. by Module, Thread, HW context…
• Identify all call stacks that called the function selected in the grid.
• Use the Timeline pane to visualize metrics over time at either the thread or platform level
  • Identify patterns, anomalies and trends in the data
  • Select a grouping level for timeline bands e.g. Process, HW context.
  • Drag and drop to select a range on the timeline area.
  • Zoom and/or filter in from the context menu
  • Filter the data in the grid or timeline view.
    • E.g. select ThreadFunction from the Thread menu to filter out other data
• Double click the function to open source and identify code lines that affect performance. Or select View Source from the context menu.
• In the Source view, see the code associated with a selection and it’s related performance metrics.
• Switch to Assembly view to see disassembled code and related performance metrics.
  • If both views are enabled, source and assembly code selections are synchronized.
• Navigate between code lines that contain the most hostspots.
• Looking for a particular object? Click the magnifier icon to open the Find panel.