Intel® Advisor User Guide

ID 766448
Date 10/31/2024
Public
Document Table of Contents

Collection Controls to Minimize Analysis Overhead

Issue

Running your target application with the Intel® Advisor can take substantially longer than running your target application without the Intel Advisor. Depending on an accuracy level and analyses you choose for a perspective, different overhead is added to your application execution time. For example:

Runtime Overhead / Analysis

Survey

Characterization

Dependencies

MAP

Target application runtime with Intel® Advisor compared to runtime without Intel® Advisor

1.1x longer

2 - 55x longer

5 - 100x longer

5 - 20x longer

Solutions

Use the following techniques to skip uninteresting parts of your target application, such as the initialization phase, and analyze only interesting parts.

Pause Collection/Resume Collection Using Annotations

Minimize collection overhead.

Applicable analyses: Survey, Dependencies.

Some analysis types recognize the structural annotations typically used in the Threading perspective workflow.

Use when:

  • Modifying/recompiling your target application is not an issue.

  • You do not want to analyze one or more uninteresting parts of your target application.

  • The interesting parts of your target application involve large workloads (because Pause/Resume API call frequency is about 1 Hz, and the operations pause and resume data collection in all processes in the analysis run, with the corresponding collection state notification to the GUI).

To pause collection, add the following annotation to your code:

  • C++: ANNOTATE_DISABLE_COLLECTION_PUSH

  • Fortran: call annotate_disable_collection_push()

  • C#: Annotate.DisableCollectionPush();

To resume collection, add the following annotation to your code:

  • C++: ANNOTATE_DISABLE_COLLECTION_POP

  • Fortran: call annotate_disable_collection_pop()

  • C#: Annotate.DisableCollectionPop();

NOTE:
C# and .NET support is deprecated starting Intel® Advisor 2021.1.

See Pause Collection and Resume Collection Annotations for more information.

Pause Collection/Resume Collection Using API Methods

Minimize collection overhead using the Intel® Instrumentation and Tracing Technology API (ITT API).

TIP:
For MPI applications, you can also use a standard MPI-specific function MPI_Pcontrol(<N>) to pause and resume data collection. This function requires no changes in the application building process, unlike the ITT API calls, which require linking of a static ITT API library. See Control Collection with an MPI_Pcontrol Function.

Applicable analyses: Survey, Characterization with Trip Counts or FLOP collection enabled.

Use when:

  • Modifying/recompiling your target application is not an issue.

  • You do not want to analyze one or more uninteresting parts of your target application.

  • The interesting parts of your target application involve large workloads (because Pause/Resume API call frequency is about 1 Hz, and the operations pause and resume data collection in all processes in the analysis run, with the corresponding collection state notification to the GUI).

Prerequisites:

  • Add the following statements to every source file you want to instrument:

    • C/C++: ittnotify.h

    • Fortran: USE ITTNOTIFY

    NOTE:
    • The ittnotify header file contains definitions of ITT API routines and important macros that provide the correct logic of API invocation from an application.

    • The ITT API is designed to incur almost zero overhead when tracing is disabled. If you need completely zero overhead, you can compile out all ITT API calls from an application by defining the INTEL_NO_ITTNOTIFY_API macro in your project at compile time, either on the compiler command line or in your source file prior to including the ittnotify header file.

  • Configure your build system to reach ITT API header file and libraries, where <install-dir> is the Intel Advisor installation directory.

    • Add the appropriate entry to your INCLUDE path:

      • C++: <install-dir>/sdk/include

      • Fortran: <install-dir>/sdk/include/lib32 or <install-dir>/sdk/include/lib64

      • Microsoft Visual Studio* IDE: Project Properties > C/C++ | Fortran > General > Additional Include Directories.

    • Add <install-dir>/sdk/lib32 or <install-dir>/sdk/lib64 to your LIBRARIES path.

      Visual Studio IDE: Project Properties > C/C++ | Fortran > General > Additional Include Libraries.

    NOTE:
    The ITT API headers, static libraries, and Fortran modules previously located at <install-dir>/include and <install-dir>/[lib32 | lib64] folders were moved to the <install-dir>/sdk/include and <install-dir>/sdk/[lib32 | lib64] folder. Copies of these files are retained at their old locations for backward compatibility and these copies should not be used for new projects.
  • Link your target application to the static library libittnotify.a (Linux* OS) or libittnotify.lib (Windows* OS) by passing -littnotify to your compiler. If tracing is enabled, this static library loads the ITT API implementation and forwards ITT API instrument data to the Intel Advisor. If tracing is disabled, the static library ignores ITT API calls, providing nearly zero instrumentation overhead.

    Visual Studio IDE: Project Properties > Linker > Input > Additional Dependencies

  • Insert _itt_pause (C/C++) or CALL ITT_PAUSE (Fortran) before uninteresting parts of your target application and the _itt_resume (C/C++) or CALL ITT_RESUME (Fortran) before interesting parts of your target application.

    Example 1: The following snippet plus the standard run control collects analysis data twice - at the beginning and the middle of the snippet:

    #include <ittnotify.h>
    int main(int argc, char* argv[])
    {
    // Do work here
    __itt_pause();
    // Do uninteresting work here
    __itt_resume();
    // Do work here
    __itt_pause();
    // Do uninteresting work here
    return 0;
    }

    Example 2: The following snippet plus the standard run control collects analysis data only once - in the middle of the snippet:

    #include <ittnotify.h>
    int main(int argc, char* argv[])
    {
    __itt_pause();
    // Do uninteresting work here
    __itt_resume();
    // Do work here
    __itt_pause();
    // Do uninteresting work here
    return 0;
    }

    Example 3: The following snippet plus the standard run control collects analysis data only once - at the end of the snippet:

    #include <ittnotify.h>
    int main(int argc, char* argv[])
    {
    __itt_pause();
    // Do uninteresting work here
    __itt_resume();
    // Do work here
    return 0;
    }

    Example 4: The following snippet plus the Start Paused control collects analysis data only once - at the end of the snippet:

    #include <ittnotify.h>
    int main(int argc, char* argv[])
    {
    // Do uninteresting work here
    __itt_resume();
    // Do work here
    return 0;
    }

After performing the prerequisites and recompiling, do one of the following:

Note: In the commands below, make sure to replace the myApplication with your application executable path and name before executing a command. If your application requires additional command line options, add them after the executable name.

  • Click the standard run control or the Start Paused control on the Analysis Workflow pane to run the desired analysis.

  • Use the advisor CLI action --collect with or without the CLI action option --start-paused to run the desired analysis. For example:

    advisor --collect=survey --project-dir=./advi_results -- ./myApplication

To attach ITT APIs to a launched application, that is, to collect API data on an application that is already launched, point the target application to the ittnotify_collector library using an environment variable:

  • Windows* OS:

    set INTEL_LIBITTNOTIFY32=<install_dir>\bin32\runtime\ittnotify_collector.dll

    set INTEL_LIBITTNOTIFY64=<install_dir>\bin64\runtime\ittnotify_collector.dll

  • Linux* OS:

    export INTEL_LIBITTNOTIFY32=<install_dir>/lib32/runtime/libittnotify_collector.so

    export INTEL_LIBITTNOTIFY64=<install_dir>/lib64/runtime/libittnotify_collector.so

NOTE:

Use the full path to the library without quotations marks.

After you complete configuration, start the instrumented application in the correct environment. Intel® Advisor collects API data even if the application is launched before the Intel® Advisor is launched.

You can find the ITT API documentation at https://github.com/intel/ittapi.

Start Target Application With Collection Paused

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled.

Use when you do not want to analyze the early phase(s) of your target application, such as the initialization phase, but you want analysis in ready mode.

Note: In the commands below, make sure to replace the myApplication with your application executable path and name before executing a command. If your application requires additional command line options, add them after the executable name.

To implement, do one of the following:

  • Click the associated control on the Analysis Workflow pane to run the desired analysis.

  • Use the advisor CLI action option --start-paused when you run the desired analysis. For example:

    advisor --collect=survey --start-paused --project-dir=./advi_results -- ./myApplication
    
NOTE:

You can use different techniques to resume collection. The most common is __itt_resume.

Start Target Application With Collection Paused/Resume Collection After N Seconds

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled.

Use when...

  • You do not want to modify/recompile your target application.

  • You do not want to analyze the initialization phase of your target application.

  • You have a good idea of the time interval of interest, but pinpointing the exact beginning of the interesting part of your target application is not important.

  • The interesting part of your target application is more than a few loops.

Note: In the commands below, make sure to replace the myApplication with your application executable path and name before executing a command. If your application requires additional command line options, add them after the executable name.

To implement, do one of the following:

  • Enable the Project Properties > Analysis Target > [Name] Analysis > Advanced > Automatically resume collection after (sec) checkbox and supply the desired value, where [Name] is Survey or Trip Counts and FLOP.

    Click the standard run control on the Analysis Workflow pane to run the desired analysis. (Collection automatically starts in the paused state.)

  • Use the advisor CLI action option --resume-after=<integer> when your run the desired analysis. For example:

    advisor --collect=survey --resume-after=30 --project-dir=./advi_results -- ./myApplication
NOTE:

Use a value representing seconds in the GUI field and milliseconds in the integer argument.

Stop Collection After N Seconds

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled, Dependencies, Memory Access Patterns.

This is the flip side of the Start target application with collection paused technique. Use when...

  • You do not want to modify/recompile your target application.

  • You do not want to analyze the end of your target application.

  • You have a good idea of the time interval of interest, but pinpointing the exact end of the interesting part of your target application is not important.

  • The interesting part of your target application is more than a few loops.

To implement, enable Project Properties > Analysis Target > [Name] Analysis > Advanced > Automatically stop collection after (sec) checkbox and supply the desired value, where [Name] is Survey or Trip Counts and FLOP.

Click the standard run control on the Analysis Workflow pane to run the desired analysis.

NOTE:

Use a value representing seconds in both the GUI field and integer argument.

Stop Collection

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled, Dependencies, Memory Access Patterns.

Use when...

  • You do not want to modify/recompile your target application.

  • You do not want to analyze the end of your target application.

  • You can detect the time interval of interest based on target application output.

  • The interesting part of your target application is more than a few loops.

To implement, do one of the following:

  • Click the associated control on the Analysis Workflow pane when running the desired analysis.

    NOTE:

    If running a Dependencies or Memory Access Patterns analysis, use the Site Coverage bar to determine when all marked loops are analyzed at least once:


  • Use the advisor CLI action --command=stop when you run the desired analysis. For example:

    advisor --command=stop --result-dir=./myAdvisorResult  
    

Manually Pause Collection/Manually Resume Collection

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled.

Use when...

  • You can detect the time interval of interest based on target application output.

  • Your need to pause or resume is unplanned and spontaneous.

To implement, do one of the following to pause analysis data collection (the target application continues running, but analysis data collection stops):

  • Click the associated control on the Analysis Workflow pane when running the desired analysis.

  • Use the advisor CLI action --command=pause when you run the desired analysis. For example:

    advisor --command=pause --result-dir=./myAdvisorResult  
    

Do one of the following to resume analysis data collection:

  • Click the associated control on the Analysis Workflow pane.

  • Use the advisor CLI action --command=resume. For example:

    advisor --command=resume --result-dir=./myAdvisorResult  
    

Attach to Process/Detach from Process

Minimize collection overhead.

Applicable analyses: Survey, Characterization with Trip Counts and FLOP collection enabled without call stacks.

This technique is similar to the Start target application with collection paused technique, except you can attach to an already running process. This is particularly beneficial if:

  • The process is a service that runs forever.

  • The launching infrastructure is relatively complicated, such as a sequence of scripts that must be modified to embed a launch collection command.

GUI:

  1. Choose Project Properties > Analysis Target > [Name] Analysis> Launch Application drop-down list > Attach to Process, where [Name] is Survey or Trip Counts and FLOP.

  2. Disable the Inherit settings from Survey Hotspots Analysis Type checkbox.

  3. Choose the Process name or PID option and identify a process.

  4. Supply other information as desired and close the Project Properties dialog box.

  5. Click the standard run control on the Analysis Workflow pane to run the desired analysis.

CLI: Use the advisor CLI action option --target-pid=<unsigned integer> or --target-process=<string> to attach to a process when running the desired analysis. For example:

advisor --collect=survey --project-dir=./advi_results --result-dir=./myAdvisorResult --target-process=myProcess 

Do one of the following to stop collecting analysis data on a process (the process continues running but analysis data collection stops):

  • Click the associated control on the Analysis Workflow pane.

  • Use the advisor CLI action --command=detach. For example:

    advisor --command=detach --result-dir=./myAdvisorResult 
    
NOTE:
  • Ensure call stacks are disabled (which is the default setting) if you run the Characterization with Trip Counts and FLOP collection enabled analysis:

    • Disable the Project Properties > Analysis Target > Trip Counts and FLOP Analysis > Advanced > Collect stacks checkbox.

    • Add the advisor CLI action option --no-stacks to the --collect command, or simply omit a --stacks action option on the --collect command.

  • Using the advisor CLI action --command=stop kills the process (which also stops analysis data collection).