extrap 4.2.3

Extra-P, automated performance modeling for HPC applications

Extra-P

Automated performance modeling for HPC applications

Extra-P is an automatic performance-modeling tool that supports the user in the identification of scalability bugs. A scalability bug is a part of the program whose scaling behavior is unintentionally poor, that is, much worse than expected. A performance model is a formula that expresses a performance metric of interest such as execution time or energy consumption as a function of one or more execution parameters such as the size of the input problem or the number of processors.

Extra-P uses measurements of various performance metrics at different execution configurations as input to generate performance models of code regions (including their calling context) as a function of the execution parameters. All it takes to search for scalability issues even in full-blown codes is to run a manageable number of small-scale performance experiments, launch Extra-P, and compare the asymptotic or extrapolated performance of the worst instances to the expectations.

Extra-P generates not only a list of potential scalability bugs but also human-readable models for all performance metrics available such as floating-point operations or bytes sent by MPI calls that can be further analyzed and compared to identify the root causes of scalability issues.

The following video on the Laboratory for Parallel Programming @ TUDa YouTube channel provides a quick introduction to Extra-P.

Extra-P is developed by TU Darmstadt – in collaboration with ETH Zurich.

For questions regarding Extra-P, please send a message to [e-mail].

---

Table of Contents

1. Requirements
2. Installation
3. How to use it
4. License
5. Citation
6. Publications

---

Requirements

• Python 3.8 or higher
• numpy
• pycubexr
• marshmallow
• packaging
• tqdm
• sklearn
• PySide6 (for GUI)
• matplotlib (for GUI)
• pyobjc-framework-Cocoa (only for GUI on macOS)

Installation

Use the following command to install Extra-P and all required packages via pip.

python -m pip install extrap --upgrade

The --upgrade forces the installation of a new version if a previous version is already installed.

Usage

Extra-P can be used in two ways, either using the command-line interface or the graphical user interface. More information about the usage of Extra-P with both interfaces can be found in the quick start guide.

Note
Extra-P is designed for weak-scaling, therefore, directly modeling of strong-scaling behaviour is not supported. Instead of modeling the runtime of your strong-scaling experiment, you can model the resource consumption, i.e., the runtime times the number of processors. Extra-P automatically offers this conversion, if it detects that strong-scaling data was loaded. If you are loading files that contain per-thread/per-rank data you should select the scaling-type upfront to run the conversion already during the import.

Graphical user interface

The graphical user interface can be started by executing the extrap-gui command.

Command line interface

The command line interface is available under the extrap command:

extrap OPTIONS (--cube | --text | --talpas | --json | --extra-p-3 | --experiment) FILEPATH

You can use different input formats as shown in the examples below:

• Text files: extrap --text test/data/text/one_parameter_1.txt
• JSON files: extrap --json test/data/json/input_1.JSON
• Talpas files: extrap --talpas test/data/talpas/talpas_1.txt
• Create model and save it to text file at the given path: extrap --out test.txt --text test/data/text/one_parameter_1.txt

You can find an overview about all command line options under docs/command-line-options.md.

License

BSD 3-Clause "New" or "Revised" License

Citation

Please cite Extra-P in your publications if it helps your research:

@inproceedings{calotoiu_ea:2013:modeling,
  author = {Calotoiu, Alexandru and Hoefler, Torsten and Poke, Marius and Wolf, Felix},
  month = {November},
  title = {Using Automated Performance Modeling to Find Scalability Bugs in Complex Codes},
  booktitle = {Proc. of the ACM/IEEE Conference on Supercomputing (SC13), Denver, CO, USA},
  year = {2013},
  pages = {1--12},
  publisher = {ACM},
  isbn = {978-1-4503-2378-9},
  doi = {10.1145/2503210.2503277}
}

Publications

1. Alexandru Calotoiu, David Beckingsale, Christopher W. Earl, Torsten Hoefler, Ian Karlin, Martin Schulz, Felix Wolf: Fast Multi-Parameter Performance Modeling. In Proc. of the 2016 IEEE International Conference on Cluster Computing ( CLUSTER), Taipei, Taiwan, pages 172–181, IEEE, September 2016. PDF

2. Marcus Ritter, Alexandru Calotoiu, Sebastian Rinke, Thorsten Reimann, Torsten Hoefler, Felix Wolf: Learning Cost-Effective Sampling Strategies for Empirical Performance Modeling. In Proc. of the 34th IEEE International Parallel and Distributed Processing Symposium (IPDPS), New Orleans, LA, USA, pages 884–895, IEEE, May 2020. PDF

3. Marcus Ritter, Alexander Geiß, Johannes Wehrstein, Alexandru Calotoiu, Thorsten Reimann, Torsten Hoefler, Felix Wolf: Noise-Resilient Empirical Performance Modeling with Deep Neural Networks. In Proc. of the 35th IEEE International Parallel and Distributed Processing Symposium (IPDPS), Portland, Oregon, USA, pages 23–34, IEEE, May 2021. PDF