Skip to content
forked from LLNL/Caliper

Caliper is an instrumentation and performance profiling library

License

Notifications You must be signed in to change notification settings

vsurjadidjaja/Caliper

 
 

Repository files navigation

Caliper: A Performance Analysis Toolbox in a Library

Github Actions Build Status Coverage

Caliper is a performance instrumentation and profiling library for HPC (high-performance computing) programs. It provides source-code annotation APIs for marking regions of interest in C, C++, and Fortran code, as well as a set of built-in performance measurement recipes for a wide range of performance engineering use cases, such as lightweight always-on profiling, event tracing, or performance monitoring. Alternatively, users can create custom measurement configurations for specialized use cases.

Caliper can either generate simple human-readable reports or machine-readable JSON or .cali files for automated data processing with user-provided scripts or analysis frameworks like Hatchet and Thicket. It can also generate detailed event traces for timeline visualizations with Perfetto and the Google Chrome trace viewer.

Features include:

  • Low-overhead source-code annotation API
  • Configuration API to control performance measurements from within an application
  • Recording program metadata for analyzing collections of runs
  • Flexible key:value data model to capture application-specific features for performance analysis
  • Fully threadsafe implementation, support for parallel programming models like MPI
  • Event-based as well as sample-based performance measurements
  • Trace and profile recording
  • Connection to third-party tools, e.g. NVidia's NSight tools, AMD ROCProf, or Intel(R) VTune(tm)
  • Measurement and profiling functionality such as timers, PAPI hardware counters, and Linux perf_events
  • Memory annotations to associate performance measurements with memory regions

Documentation

Extensive documentation is available here: https://software.llnl.gov/Caliper/

Usage examples of the C++, C, and Fortran annotation and ConfigManager APIs are provided in the examples directory.

See the "Getting started" section below for a brief tutorial.

Building and installing

You can install Caliper with the spack package manager:

$ spack install caliper

To build Caliper manually, you need cmake 3.12+ and a current C++11-compatible Compiler. Clone Caliper from github and proceed as follows:

$ git clone https://github.com/LLNL/Caliper.git
$ cd Caliper
$ mkdir build && cd build
$ cmake -DCMAKE_INSTALL_PREFIX=<path to install location> ..
$ make
$ make install

Link Caliper to a program by adding libcaliper:

$ g++ -o app app.o -L<path install location>/lib64 -lcaliper

There are many build flags to enable optional features, such as -DWITH_MPI for MPI support. See the "Build and install" section in the documentation for further information.

Getting started

Typically, we integrate Caliper into a program by marking source-code sections of interest with descriptive annotations. Performance profiling can then be enabled through the Caliper ConfigManager API or environment variables. Alternatively, third-party tools can connect to Caliper and access information provided by the source-code annotations.

Source-code annotations

Caliper's source-code annotation API allows you to mark source-code regions of interest in your program. Much of Caliper's functionality depends on these region annotations.

Caliper provides macros and functions for C, C++, and Fortran to mark functions, loops, or sections of source-code. For example, use CALI_CXX_MARK_FUNCTION to mark a function in C++:

#include <caliper/cali.h>

void foo()
{
    CALI_CXX_MARK_FUNCTION;
    // ...
}

You can mark arbitrary code regions with the CALI_MARK_BEGIN and CALI_MARK_END macros or the corresponding cali_begin_region() and cali_end_region() functions:

#include <caliper/cali.h>

// ...
CALI_MARK_BEGIN("my region");
// ...
CALI_MARK_END("my region");

The cxx-example, c-example, and fortran-example example apps show how to use Caliper in C++, C, and Fortran, respectively.

Recording performance data

With the source-code annotations in place, we can run performance measurements. By default, Caliper does not record data - we have to activate performance profiling at runtime. An easy way to do this is to use one of Caliper's built-in measurement recipes. For example, the runtime-report config prints out the time spent in the annotated regions. You can activate built-in measurement configurations with the ConfigManager API or with the CALI_CONFIG environment variable. Let's try this on Caliper's cxx-example program:

$ cd Caliper/build
$ make cxx-example
$ CALI_CONFIG=runtime-report ./examples/apps/cxx-example
Path       Min time/rank Max time/rank Avg time/rank Time %
main            0.000119      0.000119      0.000119  7.079120
  mainloop      0.000067      0.000067      0.000067  3.985723
    foo         0.000646      0.000646      0.000646 38.429506
  init          0.000017      0.000017      0.000017  1.011303

The runtime-report config works for MPI and non-MPI programs. It reports the minimum, maximum, and average exclusive time (seconds) spent in each marked code region across MPI ranks (the values are identical in non-MPI programs).

You can customize the report with additional options. Some options enable additional Caliper functionality, such as profiling MPI and CUDA functions in addition to the user-defined regions, or additional metrics like memory usage. Other measurement configurations besides runtime-report include:

  • loop-report: Print summary and time-series information for loops.
  • mpi-report: Print time spent in MPI functions.
  • callpath-sample-report: Print a time spent in functions using call-path sampling.
  • event-trace: Record a trace of region enter/exit events in .cali format.
  • hatchet-region-profile: Record a region time profile for processing with Hatchet or cali-query.

See the "Builtin configurations" section in the documentation to learn more about different configurations and their options.

You can also create entirely custom measurement configurations by selecting and configuring Caliper services manually. See the "Manual configuration" section in the documentation to learn more.

ConfigManager API

A distinctive Caliper feature is the ability to enable performance measurements programmatically with the ConfigManager API. For example, we often let users activate performance measurements with a command-line argument.

With the C++ ConfigManager API, built-in performance measurement and reporting configurations can be activated within a program using a short configuration string. This configuration string can be hard-coded in the program or provided by the user in some form, e.g. as a command-line parameter or in the programs's configuration file.

To use the ConfigManager API, create a cali::ConfigManager object, add a configuration string with add(), start the requested configuration channels with start(), and trigger output with flush():

#include <caliper/cali-manager.h>
// ...
cali::ConfigManager mgr;
mgr.add("runtime-report");
// ...
mgr.start(); // start requested performance measurement channels
// ... (program execution)
mgr.flush(); // write performance results

The cxx-example program uses the ConfigManager API to let users specify a Caliper configuration with the -P command-line argument, e.g. -P runtime-report:

$ ./examples/apps/cxx-example -P runtime-report
Path       Min time/rank Max time/rank Avg time/rank Time %
main            0.000129      0.000129      0.000129  5.952930
  mainloop      0.000080      0.000080      0.000080  3.691740
    foo         0.000719      0.000719      0.000719 33.179511
  init          0.000021      0.000021      0.000021  0.969082

See the Caliper documentation for more examples and the full API and configuration reference.

Authors

Caliper was created by David Boehme, [email protected].

A complete list of contributors is available on GitHub.

Major contributors include:

Citing Caliper

To reference Caliper in a publication, please cite the following paper:

On GitHub, you can copy this citation in APA or BibTeX format via the "Cite this repository" button. Or, see the comments in CITATION.cff for the raw BibTeX.

Release

Caliper is released under a BSD 3-clause license. See LICENSE for details.

LLNL-CODE-678900

About

Caliper is an instrumentation and performance profiling library

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 70.7%
  • C 12.2%
  • Python 8.7%
  • CMake 3.6%
  • Fortran 2.3%
  • CWeb 1.7%
  • Other 0.8%