Visible to Intel only — GUID: GUID-689276A4-9A6C-4AB9-AEC5-16DEBF3371EA
Visible to Intel only — GUID: GUID-689276A4-9A6C-4AB9-AEC5-16DEBF3371EA
Run VTune Profiler in a Container
Install a Docker* image with Intel® VTune™ Profiler and profile native or Java* applications running inside the same container or outside the container.
Prerequisites
Configure a Docker image:
Create and configure a Docker image.
For the pre-installed Intel® oneAPI Base Toolkit including VTune Profiler, you may pull an existing Docker image from the Docker Hub repository:
host> image=amr-registry.caas.intel.com/oneapi/oneapi:base-dev-ubuntu18.04 host> docker pull "$image"
To enable profiling from the container and have all host processes visible from the container, run your Docker image with --pid=host as follows:
host> docker run --pid=host --cap-add=SYS_ADMIN --cap-add=SYS_PTRACE -it "$image"
where the SYS_ADMIN value adds a capability to run hardware event-based sampling analysis; the SYS_PTRACE value enables user-mode sampling analysis.
To profile a target application running in the same container where VTune Profiler is installed, do the following:
Copy your application to the running Docker container. For example:
host> docker cp /home/samples/matrix.tar 98fec14f0c08:/var/local
where 98fec14f0c08 is your container ID.
Compile your target in the container, if required.
Install and Run VTune Profiler in a Container
These steps are NOT required if you use a Docker image with pre-installed Intel oneAPI Base Toolkit.
Install the command-line interface of VTune Profiler inside your Docker container.
Make sure to select the [2] Custom installation > [3] Change components to install and de-select components that are not required in the container environment: [3] Graphical user interface and [4] Platform Profiler.
After installation, set up environment variables for the VTune Profiler. For example, for VTune Profiler in Intel oneAPI Base Toolkit:
container> source /opt/intel/oneapi/vtune/version/env/vars.sh
Run Analysis for Your Container Target
Set up your analysis for a target running in the container, using the following supported target and analysis types:
Target Type |
Analysis Type |
---|---|
|
|
To run an analysis, enter:
vtune -collect <analysis_type> [options] -- [container_target]
For example:
container> vtune -collect hotspots -knob sampling-mode=hw -- /home/samples/matrix
Run Analysis for Your Host Target
Set up your analysis for a target running on the host, using the following supported target and analysis types:
Target Type |
Analysis Type |
---|---|
|
|
To run an analysis, enter:
vtune -collect <analysis_type> [options] -- [host_target]
For example:
container> vtune -collect hotspots -target-process java
Known Issues:
Issue: Function-level analysis is not available by default. VTune Profiler maps the samples to the binaries from user target app but it cannot resolve the functions because the binaries from the host are not available from the container.
Solution: Run the Docker container with the mounted host folder containing the binaries and specify a search directory as an argument to the vtune command.
Issue:VTune Profiler is run in the container by the root user while the app on the host is run by a non-root user. As a result, User-Mode Sampling Hotspots analysis fails to run with an error "Both target and VTune Profiler should be run by the same user".
Solution: Make sure the same user runs VTune Profiler in the container and the target app outside the container.