Limited dev time: I am currently no longer working actively on this project because I left the job where I used it extensively. If anyone wishes to take over development, that would be great, please just add a comment to issue #89.
One liner script and function submission to torque or slurm clusters with dependency tracking using python. Uses the same syntax irrespective of cluster environment!
Learn more at https://fyrd.science, https://fyrd.rtfd.io, and https://github.com/MikeDacre/fyrd
Note: Development is currently primarily on the 0.6.2 branch. The master branch reflects the highest 0.6.1 version, which is technically more stable than 0.6.2, but which has serveral bugs fixed by 0.6.2 as well as a constricted feature set. Installing via pip will install the 0.6.2 version.
| Author | Michael D Dacre <mike.dacre@gmail.com> |
| License | MIT License, property of Stanford, use as you wish |
| Version | 0.6.2b1 |
Allows simple job submission with dependency tracking and queue waiting on either torque, slurm, or locally with the multiprocessing module. It uses simple techniques to avoid overwhelming the queue and to catch bugs on the fly.
It is routinely tested on Mac OS and Linux with slurm and torque clusters, or
in the absence of a cluster, on Python versions 2.7.10, 2.7.11, 2.7.12,
3.3.0, 3.4.0, 3.5.2, 3.6.2, and 3.7-dev. The full test suite is
available in the tests folder.
Fyrd is pronounced 'feared' (sort of), it is an Anglo-Saxon term for an army, particularly an army of freemen (in this case an army of compute nodes). The logo is based on a Saxon shield commonly used by these groups. This software was formerly known as 'Python Cluster'.
For usage instructions and complete documentation see the documentation site and the fyrd_manual.pdf document in this repository.
Contents
This library was created to make working with torque or slurm clusters as easy as working with the multiprocessing library. It aims to provide:
- Easy submission of either python functions or shell scripts to torque or slurm from within python.
- Simple dependency tracking for jobs.
- The ability to submit jobs with any of the torque or slurm keyword arguments.
- Easy customization.
- Very simple usage that scales to complex applications.
- A simple queue monitoring API that functions identically with torque and slurm queues.
- A fallback local mode that allows code to run locally using the multiprocessing module without needing any changes to syntax.
To do this, all major torque and slurm keyword arguments are encoded in
dictionaries in the fyrd/options.py file using synonyms so that all arguments
are standardized on the fly. Job management is handled by the Job class in
fyrd/job.py, which accepts any of the keyword arguments in the options file.
To make submission as simple as possible, the code makes used of profiles
defined in the ~/.fyrd/profiles.txt config file. These allow simple grouping
of keyword arguments into named profiles to make submission even easier.
Dependency tracking is handled by the depends= argument to Job, which
accepts job numbers or Job objects, either singularly or as lists.
To allow simple queue management and job waiting, a Queue class is
implemented in fyrd/queue.py. It uses iterators, also defined in that file,
to parse torque or slurm queues transparently and allow access to their
attributes through the Queue class and the Queue.jobs dictionary. The Job
class uses this system to block until the job completes when either the
wait() or get() methods are called.
Note, waiting can email you when it is done, but you need to enable it in the
config file (~/.fyrd/config.txt):
[notify] mode = linux # Can be linux or smtp, linux uses the mail command notify_address = your.address@gmail.com # The following are only needed for smtp mode smtp_host = smtp.gmail.com smtp_port = 587 smtp_tls = True smtp_from = your.server@gmail.com smtp_user = None # Defaults to smtp_from # This is insecure, so use an application specific password. This should # be a read-only file with the SMTP password. After making it run: # chmod 400 ~/.fyrd/smtp_pass smtp_passfile = ~/.fyrd/smtp_pass
To allow similar functionality on a system not connected to a torque or slurm
queue, a local queue that behaves similarly, including allowing dependency
tracking, is implemented in the fyrd/jobqueue.py file. It is based on
multiprocessing but behaves like torque. It is not a good idea to use this
module in place of multiprocessing due to the dependency tracking overhead, it
is primarily intended as a fallback, but it does work well enough to use
independently. Note: the local mode currently is quite slow, as the overhead
for job management means that 100% of each available CPU is not used, only
around 80% is. The local mode still works fine as a fallback or for testing
code, but it is important to remember that fyrd is meant primarily for large
cluster use.
As all clusters are different, common alterable parameters are defined in a
config file located at ~/.fyrd/config.txt. This includes an option for max
queue size, which makes job submission block until the queue has opened up,
preventing job submission failure on systems with queue limits (most clusters).
To make life easier, a bunch of simple wrapper functions are defined in
fyrd/basic.py that allow submission without having to worry about using the
class system, or to submit existing job files. Several helper function are also
created in fyrd/helpers.py that allow the automation of more complex tasks,
like running apply on a pandas dataframe in parallel on the cluster
(fyrd.helpers.parapply()).
The end result is that submitting 10 thousand very small jobs to a small cluster can be done like this:
jobs = []
for i in huge_list:
jobs.append(fyrd.Job(my_function, (i,), profile='small').submit())
results = fyrd.get(jobs)The results list in this example will contain the function outputs, even if those outputs are integers, objects, or other Python types. Similarly, shell scripts can be run like this:
script = r"""zcat {} | grep "#config" | awk '{{split($1,a,"."); print a[2]"\t"$2}}'"""
jobs = []
for i in [i for i in os.listdir('.') if i.endswith('.gz')]:
jobs.append(fyrd.Job(script.format(i), profile='long').submit())
results = fyrd.get(jobs)
for i in results:
print(i.stdout)Results will contain the contents of STDOUT for the submitted script
Here is the same code with dependency tracking:
script = r"""zcat {} | grep "#config" | awk '{{split($1,a,"."); print a[2]"\t"$2}}'"""
jobs = []
jobs2 = []
for i in [i for i in os.listdir('.') if i.endswith('.gz')]:
j = fyrd.Job(script.format(i), profile='long').submit()
jobs.append(j)
jobs2.append(fyrd.Job(my_function, depends=j).submit())
results = []
for i in jobs2:
i.wait()
results.append(i.out)As you can see, the profile keyword is not required, if not supplied the
default profile is used. It is also important to note that .out will contain
the same contents as .stdout for all script submissions, but for function
submissions, .out contains the function output, not STDOUT.
Note, to submit simple functions, I recommend that you use the jobify
decorator instead:
>>> import fyrd
>>> @fyrd.jobify(name='test_job', mem='1GB')
... def test(string, iterations=4):
... """This does basically nothing!"""
... outstring = ""
... for i in range(iterations):
... outstring += "Version {0}: {1}".format(i, string)
... return outstring
...
>>> test?
Signature: test(*args, **kwargs)
Docstring:
This is a fyrd.job.Job decorated function.
When you call it it will return a Job object from which you can get
the results with the ``.get()`` method.
Original Docstring:
This does basically nothing!
File: ~/code/fyrd/fyrd/helpers.py
Type: function
>>> j = test('hi')
>>> j.get()
'Version 0: hiVersion 1: hiVersion 2: hiVersion 3: hi'Fyrd provides a few command line tools to make little jobs easier. The main
tool is fyrd. Running fyrd --help will give instructions on use, something
like this:
usage: fyrd [-h] [-v] [-V]
{run,submit,wait,queue,conf,prof,keywords,clean,local} ...
Manage fyrd config, profiles, and queue.
============ ======================================
Author Michael D Dacre <mike.dacre@gmail.com>
Organization Stanford University
License MIT License, use as you wish
Version 0.6.2-beta1
============ ======================================
positional arguments:
{run,submit,wait,queue,conf,prof,keywords,clean,local}
run (r) Run simple shell scripts
submit (sub, s) Submit existing job files
wait (w) Wait for jobs
queue (q) Search the queue
conf (config) View and manage the config
prof (profile) Manage profiles
keywords (keys, options)
Print available keyword arguments.
clean Clean up a job directory
local (server) Manage the local queue server
optional arguments:
-h, --help show this help message and exit
-v, --verbose Show debug outputs
-V, --version Print version string
The keywords each have their own help menus and are fairly self-explanatory.
The conf and profile arguments allow you to edit the fyrd config and
cluster profiles without having to directly edit the config files in the
~/.fyrd/ directory.
The keywords argument is a help function only, it prints all possible keyword
arguments that can be used in cluster submissions.
queue allows you to query the queue in the same way that squeue or qstat
would, with a few extra functions to make it easy to see only your jobs, or
only your running jobs.
There is another command line tool provided myqueue or myq (both are the
same), these tools are just wrappers for fyrd queue and they make it really
fast to query a torque or slurm queue on any machine. e.g. myq -r will show
you all your currently running jobs, myq -r -c will display a count of all
currently running jobs, and myq -r -l will dump a list of job numbers only to
the console, really useful when combined with xargs, e.g. myq -r -l | xargs
qdel.
The wait command just blocks until the provided job numbers complete, and
can send you an email when it completes, see the config info above.
And the clean command provides options to clean out a job directory that
contains leftover files from a fyrd session.
This module will work with Python 2.7+ on Linux and Mac OS systems.
The betas are on PyPI, and can be installed directly from there:
pip install fyrd
fyrd conf initTo install a specific tag from github, do the following:
pip install https://github.com/MikeDacre/fyrd/archive/v0.6.1b9.tar.gz
fyrd conf initTo get the latest version:
pip install https://github.com/MikeDacre/fyrd/tarball/master
fyrd conf initTo get the development version (still pretty stable):
pip install https://github.com/MikeDacre/fyrd/tarball/dev
fyrd conf initThe fyrd conf init command initializes your environment interactively by
asking questions about the local cluster system.
I recommend installing using anaconda or pyenv, this will make your life much simpler, but is not required.
In general you want either pyenv or user
level install (pip install --user) even if you have sudo access, as most
cluster environments share /home/<user> across the cluster, making this module
available everywhere. Anaconda will work if it is installed in a cross-cluster
capacity, usually as a module (with lmod, e.g. module load anaconda). An
install to the system python will usually fail as cluster nodes need to have
access to the module also.
Importing is simple:
import fyrdThis software requires the following external modules:
- dill — which makes function submission more stable
- tabulate — allows readable printing of help
- six — makes python2/3 cross-compatibility easier
- tblib — allows me to pass Tracebacks between nodes
- tqdm — pretty progress bars for multi-job get and wait
- sqlalchemy — used in local mode to track jobs
- Pyro4 — used in local mode to make a daemon
In order to submit functions to the cluster, this module must import them on the compute node. This means that all of your python modules must be available on every compute node.
By default, the same python executable used for submission is used on the cluster to run functions, however, this can be overridden by the 'generic_python' option on the cluster. If using this option, you must install all of your local modules on the cluster also.
To avoid pain and debugging, you can do this manually by running this on your login node:
freeze --local | grep -v '^\-e' | cut -d = -f 1 > module_list.txtAnd then on the compute nodes:
cat module_list.txt | xargs pip install --userAlternately, if your pyenv is available on the cluster nodes, then all of your modules are already available, so you don't need to worry about this!
To fully test this software, I use py.test tests written in the tests folder.
Unfortunately, local queue tests do not work with py.test, so I have separated
them out into the local_queue.py script. To run all tests, run python
tests/run_tests.py.
To ensure sensible testing always, I use buildkite, which is an amazing piece of software. It integrates into this repository and runs tests on all python versions I support on my two clusters (a slurm cluster and a torque cluster) every day and on every push or pull request. I also use travis ci to run local queue tests, and codacy to monitor code style.
All code in the master branch must pass the travis-ci and buildkite tests, code in dev also usually passes those test, but it is not guaranteed. All other branches are unstable and will often fail the tests.
I use the following work-flow to release versions of fyrd:
- Develop new features and fix new bugs in a feature branch
- Write tests for the new feature
- When all tests are passing, merge into dev
- Do more extensive manual testing in dev, possibly add additional commits.
- Repeat the above for other related features and bugs
- When a related set of fixes and features are done and well tested, merge into master with a pull request through github, all travis and buildkite tests must pass for the merge to work.
- At some point after the new features are in master, add a new tagged beta release.
- After the beta is determined to be stable and all issues attached to that version milestone are resolved, create a non-beta tag
New releases are added when enough features and fixes have accumulated to justify it, new minor version are added only when there are very large changes in the code and are always tracked by milestones.
While this project is still in its infancy, the API cannot be considered stable and the major version will remain 0. once version 1.0 is reached, any API changes will result in a major version change.
As such, and non-beta release can be considered stable, beta releases and the master branch are very likely to be stable, dev is usually but not always stable, all other branches are very unstable.
If you have any trouble with this software add an issue in https://github.com/MikeDacre/fyrd/issues
For peculiar technical questions or help getting the code installed, email me at mike.dacre@gmail.com.
I am always looking for help with this software, and I will gladly accept pull requests. In particular, I am looking for help with:
- Testing the code in different cluster environments
- Expanding the list of keyword options
- Adding new clusters other than torque and slurm
- Implementing new features in the issues section
If you are interested in helping out with any of those things, or if you would be willing to give me access to your cluster to allow me to run tests and port fyrd to your environment, please contact me.
If you are planning on contributing and submitting a pull request, please follow these rules:
- Follow the code style as closely as possible, I am a little obsessive about that
- If you add new functions or features: - Add some tests to the test suite that fully test your new feature - Add notes to the documentation on what your feature does and how it works
- Make sure your code passes the full test suite, which means you need to run
python tests/run_tests.pyfrom the root of the repository at a bare minimum. Ideally, you will install pyenv and runbash tests/pyenv_tests.py - Squash all of your commits into a single commit with a well written and informative commit message.
- Send me a pull request to either the
devormasterbranches.
It may take a few days for me to fully review your pull request, as I will test
it extensively. If it is a big new feature implementation I may request that
you send the pull request to the dev branch instead of to master.
I gave this project the name 'Fyrd' in honour of my grandmother, Hélène Sandolphen, who was a scholar of old English. It is the old Anglo-Saxon word for 'army', and this code gives you an army of workers on any machine so it seemed appropriate.
The project used to be called "Python Cluster", which is more descriptive but frankly boring. Also, about half a dozen other projects have almost the same name, so it made no sense to keep that name and put the project onto PyPI.
This software is much more powerful that this document gives it credit for, to get the most out of it, read the docs at https://fyrd.readthedocs.org or get the PDF version from the file in docs/fyrd_manual.pdf.