Application Definition Developers Guide

Application definition files represent building blocks to create experiments from. They are intended to be portable, and as a result should not contain any system specific logic. Each definition is a python module, with a specifically named python class contained inside. The python class can be written with provided directives, or can optionally override internal functions for more advanced behavior.

This gudie will provide general steps for creating a new application definition file.

Preparation

To begin, it can be useful to prepare a few things about the application you are writing a definition file for. These include:

  1. Instructions for compiling / installing your application

  2. A set of input files you would like to create workloads for

  3. A working set of execution commands

  4. Information about the output from the application

Most of these steps require some research, or up front exploration of the application, but are incredibly helpful in lowering the effort required to write an application definition file.

Below we will provide some basics of how to get started with these steps. However, this guide will not provide exhaustive information for all applications. So, in general, this step is left to you to complete.

Compilation / Installation

Ramble has first party support for some package managers. The provided package managers can be seen through the ramble list --type package_managers command. To improve provenance information that Ramble is able to track, we strongly recommend using a supported package manager.

You have three options at this stage. The first (and arguably more complicated) option is to add support for a new package manager, if your application is not included in any of the supported package manager yet. The second is to add your application to one of the supported package managers. The third and final option is to manually install your application and use the default setting for a package manager which disables package manager support.

This might include writing a definition file for the targeted package manager. This guide will not walk you through this process, however the actual steps vary dramatically from one package manager to another. Spack has documentation to help write package definition files.

Collecting Inputs

Some, but not all, applications will require input files. They may require multiple input files per experiment, or a single input file for each experiment. Applications can also have a wide range of input files, that each represent different workloads.

In Ramble, a workload can have an arbitrary number of input files. Before writing an application definition file, it is useful to collect, and organize the application’s input files based on the workloads you want to create. This might only involve collecting URLs and SHA256 checksums for the input files.

Testing Execution

Since an application definition file will be used to generate experiments, it is necessary to understand how to execute an experiment with the application. For this step, it can be useful to manually test the application with a set of input files for a specific workload.

Output Information

As a manual execution is recommended, once a successful test has been performed, it can be useful to retain the output of the application. Examining this output, you can determine what figures of merit might be useful to extract from the application output.

However, this step can be performed after experiments are functional from the application definition file as well.

Application Definition Creation

Application definition files are stored within object repositories. These repositories generally store all applications within a directory named applications, however each repository can control this through their own config file ( repo.yaml ).

Within the repository, each application definition file is a python module that is stored within a directory named for the application. As an example, Ramble comes with a repository named builtin. This repository contains several standard application definitions that are provided to the community. One of the application definition files provided is HPL. The HPL application definition file is named application.py and is stored within a directory named hpl. Within the application.py file, a python class is defined with a similar name to the application directory. Ramble’s application definition naming syntax follows Spack’s package naming rules.

Base Classes

Ramble provides base classes which can be inherited from when creating new application definition files. Currently, these are used to abstract the package manager logic, but more generally change the behavior of the underlying application definitions. These can be seen in more detail in ramble.application_types.

New application definitions can also inherit their behavior from other application classes to replicate aspects of their behavior.

Existing application classes can be referenced using the: from ramble.app.builtin.<application_name> import <application_class> syntax.

Writing an application definition

After an application’s application.py file is created, Ramble’s language features can be used to fill out the application definition. These language features provide directives which define specific portions of the application’s functionality. This guide will introduce some of the basic language features to create functional application definition files, but will not be exhaustive. For an exhaustive list of application language features, see ramble.language.

The directives from Ramble’s application language are placed alongside class variables, as in:

class Hpl(ExecutableApplication):
    executable(....)
    executable(....)
    input_file(....)
    input_file(....)
    workload(....)

Executables

A named executable in Ramble is one or more commands that should be executed together within an experiment. Ramble contains a directive for defining named executables ramble.language.application_language.executable()

Having performed a test execution in Testing Execution, you should be able to transcribe the execution commands into executable statements.

It is important to make sure every step needed to go from a vanilla input file to performing an experiment is captured in named executables though.

Input Files

A named input file in Ramble describes a URL, a SHA256 checksum, and some additional information about a file that a workload will require for its experiments. The input_file directive (ramble.language.application_language.input_file()) can be used to define a named input file in an application definition file.

After collecting the input files needed for the workloads you are defining (as in Collecting Inputs), each input file can be written as its own input_file directive.

Workloads

Having used executable and input_file directives, these can now be pieced together into a workload, using the workload directive (ramble.language.application_language.workload()).

The workload directive is used to define a named workload from which experiments can be generated. A workload in Ramble is defined as the pairing of one or more named executables with zero or more named input files. Defining a workload in an application.py allows it to be used within a Workspace Configuration File and will be shown when executing ramble info <app> on the named application.

Workload Variables

While a workload by itself can generate an experiment, sometimes a variable should be exposed that can allow a parameter study or help abstract the definition of the workload (such as executable commands). Each workload can have an arbitrary number of workload variables, defined by ramble.language.application_language.workload_variable().

Each variable has a default value, which can be override within a Workspace Configuration File.

Success Criteria

Success criteria help Ramble identify if an experiment was executed successfully or not. This information is extracted when ramble workspace analyze is executed to help convey if the extract figures of merti should be considered valid or not.

Applications can define any number of named success criteria, using ramble.language.shared_language.success_criteria(). For an experiment to be considered successful, all of its success criteria must resolve to True.

The simplest success criteria is a basic string match that requires a specific string show up in an experiment’s output file. More complex success criteria can also be written (including defining an evaluate_success function within the application definition file).

Figures Of Merit

Named figures of merit represent quantities that ramble should extract from an experiment. They are allowed to relate to any metric of interest, whether it is a physical quantity (such as total mass or energy), or a performance quantity (such as wallclock time), or some other application output.

Each figure of merit is defined by ramble.language.shared_language.figure_of_merit() and contains information about where the metric can be found, what the units of the metric are, and how to extract it from a given output file.

Figure Of Merit Contexts

Sometimes, a figure of merit needs additional information (such as what timestep it was collected from). To augment a figure of merit with this additional information, Ramble has the ability to define a figure of merit context. Each context represents a grouping of figures of merit that are collected together. A figure of merit context can be defined using ramble.language.shared_language.figure_of_merit_context().

File path handling

Ramble provides a utility function ramble.util.file_util.get_file_path() that should be used when referencing file paths in application definitions. This helps with Ramble to properly mock out these paths during unit testing, where the files may not exist under the dry-run setting.

Package Manager Directives

Each package manager will be its own base class, but additionally there are directives that are intended to be package manager specific. As an example, there are directives for Spack defined by:

These provide Ramble with information about how Spack could install and require packages. For more information, see the above reference.

Usage While Developing

It can be useful to test an application.py while developing it, to make sure it behaves as expected. This section will describe how you can interact with the various parts of an application definition file. This section will provide you with tips to help accelerate development and testing of an application development file

Generating Experiments

The most useful part of an application definition file is the ability to generate new experiments for its workloads. To do this, the application definition needs to contain a complete definition of at least one workload. This includes its executables, input files, and workload variables.

Once this is complete, a workspace can be configured (following Workspace Configuration File) to create experiments from the new workload. After setting up the workspace, requested experiments directories will be created following Workspace Structure. In order to debug any issues with the experiments, you can use the dry-run option from Setting up a Workspace. Additionally, you can filter the experiments you want to setup using the --where option, as in Filtering Experiments

Analyzing Results

Experiment analysis only works once figures of merit, and success criteria are defined. Without these, Ramble has no information about how to extract relevant metrics.

Once an experiment can be executed using the application.py file, you can analyze the experiment to extract all of the figures of merit. However, if you have the output file from Output Information, you can copy it into one of the experiment directories to allow analyze to extract the correct information without having to execute the experiment.