Tutorial

This tutorial is designed to step new users through the basics of setting up a signac data space, defining and executing a simple workflow, and analyzing the data. For the complete code corresponding to this tutorial, see the Ideal Gas example.

Basics

Initializing the data space

In this tutorial, we will perform a simple study of the pressure-volume (p-V) relationship of a noble gas. As a first approximation, we could model the gas as an ideal gas, so the ideal gas law applies:

\[p V = N k_B T\]

Therefore, we can assume that the volume \(V\) can be directly calculated as a function of system size \(N\), Boltzmann’s constant \(k_B\), and temperature \(T\).

To test this relationship, we start by creating an empty project directory where we will place all the code and data associated with this computational study.

~ $ mkdir ideal_gas_project
~ $ cd ideal_gas_project/
~/ideal_gas_project $

We then proceed by initializing the data space within a Python script called init.py:

# init.py
import signac

project = signac.init_project('ideal-gas-project')

for p in range(1, 10):
    sp = {'p': p, 'kT': 1.0, 'N': 1000}
    job = project.open_job(sp)
    job.init()

The signac.init_project() function initializes the signac project in the current working directory by creating a configuration file called signac.rc. The location of this file defines the project root directory. We can access the project interface from anywhere within and below the root directory by calling the signac.get_project() function, or from outside this directory by providing an explicit path, e.g., signac.get_project('~/ideal_gas_project').

Note

The name of the project stored in the configuration file is independent of the directory name it resides in.

We can verify that the initialization worked by examining the implicit schema of the project we just created:

~/ideal_gas_project $ signac schema
{
 'N': 'int([1000], 1)',
 'kT': 'float([1.0], 1)',
 'p': 'int([1, 2, 3, ..., 9, 10], 10)',
}

The output of the $ signac schema command gives us a brief overview of all keys that were used as well as their value (range).

Note

The job.init() function is idempotent, meaning that it is safe to call it multiple times even after a job has already been initialized. It is good practice make all steps that are part of the data space initialization routine idempotent.

Exploring the data space

The core function that signac offers is the ability to associate metadata — for example, a specific set of parameters such as temperature, pressure, and system size — with a distinct directory on the file system that contains all data related to said metadata. The open_job() method associates the metadata specified as its first argument with a distinct directory called a job workspace. These directories are located in the workspace sub-directory within the project directory and the directory name is the so called job id.

~/ideal_gas_project $ ls -1 workspace/
03585df0f87fada67bd0f540c102cce7
22a51374466c4e01ef0e67e65f73c52e
71855b321a04dd9ee27ce6c9cc0436f4
# ...

The job id is a highly compact, unambiguous representation of the full metadata, i.e., a distinct set of key-value pairs will always map to the same job id. However, it can also be somewhat cryptic, especially for users who would like to browse the data directly on the file system. Fortunately, you don’t need to worry about this internal representation of the data space while you are actively working with the data. Instead, you can create a linked view with the signac view command:

~/ideal_gas_project $ signac view
~/ideal_gas_project $ ls -d view/p/*
view/p/1  view/p/2  view/p/3  view/p/4  view/p/5  view/p/6  view/p/7  view/p/8  view/p/9

The linked view is the most compact representation of the data space in form of a nested directory structure. Most compact means in this case, that signac detected that the values for kT and N are constant across all jobs and are therefore safely omitted. It is designed to provide a human-readable representation of the metadata in the form of a nested directory structure. Each directory contains a job directory, which is a symbolic link to the actual workspace directory.

Note

Make sure to update the view paths by executing the $ signac view command (or equivalently with the create_linked_view() method) everytime you add or remove jobs from your data space.

Interacting with the signac project

You interact with the signac project on the command line using the signac command. You can also interact with the project within Python via the signac.Project class. You can obtain an instance of that class within the project root directory and all sub-directories with:

>>> import signac
>>> project = signac.get_project()

Iterating through all jobs within the data space is then as easy as:

>>> for job in project:
...     print(job)
...
03585df0f87fada67bd0f540c102cce7
22a51374466c4e01ef0e67e65f73c52e
71855b321a04dd9ee27ce6c9cc0436f4
# ...

We can iterate through a select set of jobs with the find_jobs() method in combination with a query expression:

>>> for job in project.find_jobs({"kT": 1.0, "p.$lt": 3.0}):
...     print(job, job.sp.p)
...
742c883cbee8e417bbb236d40aea9543 1
ee550647e3f707b251eeb094f43d434c 2
>>>

In this example we selected all jobs, where the value for \(kT\) is equal to 1.0 – which would be all of them – and where the value for \(p\) is less than 3.0. The equivalent selection on the command line would be achieved with $ signac find kT 1.0 p.\$lt 3.0. See the detailed Query API documentation for more information on how to find and select specific jobs.

Note

The following expressions are all equivalent: for job in project:, for job in project.find_jobs():, and for job in project.find_jobs(None):.

Operating on the data space

Each job represents a data set associated with specific metadata. The point is to generate data which is a function of that metadata. Within the framework’s language, such a function is called a data space operation.

Coming back to our example, we could implement a very simple operation that calculates the volume \(V\) as a function of our metadata like this:

def volume(N, kT, p):
    return N * kT / p

Let’s store the volume within our data space in a file called volume.txt, by implementing this function in a Python script called project.py:

# project.py
import signac


def compute_volume(job):
    volume = job.sp.N * job.sp.kT / job.sp.p
    with open(job.fn('volume.txt'), 'w') as file:
        file.write(str(volume) + '\n')

project = signac.get_project()
for job in project:
    compute_volume(job)

Executing this script will calculate and store the volume for each pressure-temperature combination in a file called volume.txt within each job’s workspace.

Note

The job.fn('volume.txt') expression is a short-cut for os.path.join(job.workspace(), 'volume.txt').

Workflows

Implementing a simple workflow

In many cases, it is desirable to avoid the repeat execution of data space operations, especially if they are not idempotent or are significantly more expensive than our simple example. For this, we will incorporate the compute_volume() function into a workflow using the FlowProject class. We slightly modify our project.py script:

# project.py
from flow import FlowProject


@FlowProject.operation
def compute_volume(job):
    volume = job.sp.N * job.sp.kT / job.sp.p
    with open(job.fn('volume.txt'), 'w') as file:
        file.write(str(volume) + '\n')


if __name__ == '__main__':
    FlowProject().main()

The operation() decorator identifies the compute_volume function as an operation function of our project. Furthermore, it is now directly executable from the command line via an interface provided by the main() method.

We can then execute all operations defined within the project with:

~/ideal_gas_project $ python project.py run
Execute operation 'compute_volume(03585df0f87fada67bd0f540c102cce7)'...
Execute operation 'compute_volume(22a51374466c4e01ef0e67e65f73c52e)'...
Execute operation 'compute_volume(71855b321a04dd9ee27ce6c9cc0436f4)'...
# ...

However, if you execute this in your own terminal, you might have noticed a bunch of warning messages printed out at the end, that read similar to:

Operation 'compute_volume(03585df0f87fada67bd0f540c102cce7)' exceeds max. # of allowed passes (1).
Operation 'compute_volume(22a51374466c4e01ef0e67e65f73c52e)' exceeds max. # of allowed passes (1).
# and so on

That is because by default, the run command will continue to execute all defined operations until they are considered completed. An operation is considered completed when all its post conditions are met, and it is up to the user to define those post conditions. Since we have not defined any post conditions yet, signac would continue to execute the same operation indefinitely.

For this example, a good post condition would be the existence of the volume.txt file. To tell the FlowProject class when an operation is completed, we can modify the above example by adding a function that defines this condition:

# project.py
from flow import FlowProject


def volume_computed(job):
    return os.path.isfile("volume.txt")


@FlowProject.operation
@FlowProject.post(volume_computed)
def compute_volume(job):
    volume = job.sp.N * job.sp.kT / job.sp.p
    with open(job.fn('volume.txt'), 'w') as file:
        file.write(str(volume) + '\n')


if __name__ == '__main__':
    FlowProject().main()

Tip

Simple conditions can be conveniently defined inline as lambda expressions: @FlowProject.post(lambda job: job.isfile("volume.txt")).

We can check that we implemented the condition correctly by executing $ python project.py run again. This should now return without any message because all operations have already been completed.

Note

To simply, execute a specific operation from the command line ignoring all logic, use the exec command, e.g.: $ python project.py exec compute_volume. This command (as well as the run command) also accepts jobs as arguments, so you can specify that you only want to run operations for a specific set of jobs.

Extending the workflow

So far we learned how to define and implement data space operations and how to define simple post conditions to control the execution of said operations. In the next step, we will learn how to integrate multiple operations into a cohesive workflow.

First, let’s verify that the volume has actually been computed for all jobs. For this we transform the volume_computed() function into a label function by decorating it with the label() decorator:

# project.py
from flow import FlowProject


@FlowProject.label
def volume_computed(job):
    return job.isfile("volume.txt")

# ...

We can then view the project’s status with the status command:

~/ideal_gas_project $ python project.py status
Generate output...

Status project 'ideal-gas-project':
Total # of jobs: 10

label            progress
---------------  --------------------------------------------------
volume_computed  |########################################| 100.00%

That means that there is a volume.txt file in each and every job workspace directory.

Let’s assume that instead of storing the volume in a text file, we wanted to store in it in a JSON file called data.json. Since we are pretending that computing the volume is an expensive operation, we will implement a second operation that copies the result stored in the volume.txt file into the data.json file instead of recomputing it:

# project.py
from flow import FlowProject
import json
# ...

@FlowProject.operation
@FlowProject.pre(volume_computed)
@FlowProject.post.isfile("data.json")
def store_volume_in_json_file(job):
    with open(job.fn("volume.txt")) as textfile:
        with open(job.fn("data.json"), "w") as jsonfile:
            data = {"volume": float(textfile.read())}
            jsonfile.write(json.dumps(data) + "\n")

# ...

Here we reused the volume_computed condition function as a pre-condition and took advantage of the post.isfile short-cut function to define the post-condition for this operation function.

Important

An operation function is eligible for execution if all pre-conditions are met, at least one post-condition is not met and the operation is not currently submitted or running.

Next, instead of running this new function for all jobs, let’s test it for one job first.

~/ideal_gas_project $ python project.py run -n 1
Execute operation 'store_volume_in_json_file(742c883cbee8e417bbb236d40aea9543)'...

We can verify the output with:

~/ideal_gas_project $ cat workspace/742c883cbee8e417bbb236d40aea9543/data.json
{"volume": 1000.0}

Since that seems right, we can then store all other volumes in the respective data.json files by executing $ python project run.

Tip

We could further simplify our workflow definition by replacing the pre(volume_computed) condition with pre.after(compute_volume), which is a short-cut to reuse all of compute_volume()’s post-conditions as pre-conditions for the store_volume_in_json_file() operation.

The job document

Storing results in JSON format – as shown in the previous section – is good practice because the JSON format is an open, human-readable format, and parsers are readily available in a wide range of languages. Because of this, signac stores all metadata in JSON files and in addition comes with a built-in JSON-storage container for each job (see: The Job Document).

Let’s add another operation to our project.py script that stores the volume in the job document:

# project.py
# ...

@FlowProject.operation
@FlowProject.pre.after(compute_volume)
@FlowProject.post(lambda job: 'volume' in job.document)
def store_volume_in_document(job):
    with open(job.fn("volume.txt")) as textfile:
        job.document.volume = float(textfile.read())

Besides needing fewer lines of code, storing data in the job document has one more distinct advantage: it is directly searchable. That means that we can find and select jobs based on its content.

Executing the $ python project.py run command after adding the above function to the project.py script will store all volume in the job documents. We can then inspect all searchable data with the $ signac find command in combination with the --show option:

~/ideal_gas_project $ signac find --show
03585df0f87fada67bd0f540c102cce7
{'N': 1000, 'kT': 1.0, 'p': 3}
{'volume': 333.3333333333333}
22a51374466c4e01ef0e67e65f73c52e
{'N': 1000, 'kT': 1.0, 'p': 5}
{'volume': 200.0}
71855b321a04dd9ee27ce6c9cc0436f4
{'N': 1000, 'kT': 1.0, 'p': 4}
{'volume': 250.0}
# ...

When executed with --show, the find command not only prints the job id, but also the metadata and the document for each job. In addition to selecting by metadata as shown earlier, we can also find and select jobs by their job document content, e.g.:

~/ideal_gas_project $ signac find --doc-filter volume.\$lte 125 --show
Interpreted filter arguments as '{"volume.$lte": 125}'.
df1794892c1ec0909e5955079754fb0b
{'N': 1000, 'kT': 1.0, 'p': 10}
{'volume': 100.0}
dbe8094b72da6b3dd7c8f17abdcb7608
{'N': 1000, 'kT': 1.0, 'p': 9}
{'volume': 111.11111111111111}
97ac0114bb2269561556b16aef030d43
{'N': 1000, 'kT': 1.0, 'p': 8}
{'volume': 125.0}

Note

The job document is a feature of the core signac package, and can be used even outside the context of a FlowProject.

Job scripts and cluster submission

Generating scripts

So far, we executed all operations directly on the command line with the run command. However we can also generate scripts for execution, which is especially relevant if you intend to submit the workflow to a scheduling system typically encountered in high-performance computing (HPC) environments.

Scripts are generated using the jinja2 templating system, but you don’t have to worry about that unless you want to change any of the default templates.

We can generate a script for the execution of the next eligible operations with the script command. We need to reset our workflow before we can test that:

~/ideal_gas_project $ rm -r workspace/
~/ideal_gas_project $ python init.py

Let’s start by generating a script for the execution of up to two eligible operations:

~/ideal_gas_project $ python project.py script -n 2
set -e
set -u

cd /Users/csadorf/ideal_gas_project

# Operation 'compute_volume' for job '03585df0f87fada67bd0f540c102cce7':
python project.py exec compute_volume 03585df0f87fada67bd0f540c102cce7
# Operation 'compute_volume' for job '22a51374466c4e01ef0e67e65f73c52e':
python project.py exec compute_volume 22a51374466c4e01ef0e67e65f73c52e

By default, the generated script will change into the project root directory and then execute the command for each next eligible operation for all selected jobs. We then have two ways to run this script. One option would be to pipe it into a file and then execute it:

~/ideal_gas_project $ python project.py script > run.sh
~/ideal_gas_project $ /bin/bash run.sh

Alternatively, we could pipe it directly into the command processor:

~/ideal_gas_project $ python project.py script | /bin/bash

Executing the script command again, we see that it would now execute both the store_volume_in_document and the store_volume_in_json_file operation, since both share the same pre-conditions:

~/ideal_gas_project $ python project.py script -n 2
set -e
set -u

cd /Users/csadorf/ideal_gas_project

# Operation 'store_volume_in_document' for job '03585df0f87fada67bd0f540c102cce7':
python project.py exec store_volume_in_document 03585df0f87fada67bd0f540c102cce7
# Operation 'store_volume_in_json_file' for job '03585df0f87fada67bd0f540c102cce7':
python project.py exec store_volume_in_json_file 03585df0f87fada67bd0f540c102cce7

If we wanted to customize the script generation, we could either extend the base template or simply replace the default template with our own. To replace the default template, we can put a template script called script.sh into a directory called templates within the project root directory. A simple template script might look like this:

cd {{ project.config.project_dir }}

{% for operation in operations %}
{{ operation.cmd }}
{% endfor %}

Storing the above template within a file called templates/script.sh will now change the output of the script command to:

~/ideal_gas_project $ python project.py script -n 2
cd /Users/csadorf/ideal_gas_project

python project.py exec store_volume_in_document 03585df0f87fada67bd0f540c102cce7
python project.py exec store_volume_in_json_file 03585df0f87fada67bd0f540c102cce7

Please see $ python project.py script --template-help to get more information on how to write and use custom templates.

Submit operations to a scheduling system

In addition to executing operations directly on the command line and generating scripts, signac can also submit operations to a scheduler such as SLURM. This is essentially equivalent to generating a script as described in the previous section, but in this case the script will also contain the relevant scheduler directives such as the number of processors to request. In addition, signac will also keep track of submitted operations in addition to workflow progress, which almost completely automates the submission process as well as preventing the accidental repeated submission of operations.

To use this feature, make sure that you are on a system with any of the supported schedulers and then run the $ python project.py submit command.