Cloudmesh Controlled Computing through Workflows
Gregor von Laszewski (laszewski@gmail.com)$^*$, Jacques Fleischer
$^*$ Corresponding author
@misc{las-2022-hybrid-cc,
title = {Hybrid Reusable Computational Analytics Workflow
Management with Cloudmesh},
author = {Gregor von Laszewski and J. P. Fleischer and
Geoffrey C. Fox},
year = 2022,
eprint = {2210.16941},
archivePrefix ={arXiv},
primaryClass = {cs.DC},
url = {https://arxiv.org/pdf/2210.16941},
urlOPT =
{https://github.com/cyberaide/paper-cloudmesh-cc/raw/main/vonLaszewski-cloudmesh-cc.pdf}
}
High-performance computing (HPC) is for decades a very important tool for science. Scientific tasks can be leveraging the processing power of a supercomputer so they can run at previously unobtainable high speeds or utilize specialized hardware for acceleration that otherwise are not available to the user. HPC can be used for analytic programs that leverage machine learning applied to large data sets to, for example, predict future values or to model current states. For such high-complexity projects, there are often multiple complex programs that may be running repeatedly in either competition or cooperation. Leveraging for example computational GPUs leads to several times higher performance when applied to deep learning algorithms. With such projects, program execution is submitted as a job to a typically remote HPC center, where time is billed as node hours. Such projects must have a service that lets the user manage and execute without supervision. We have created a service that lets the user run jobs across multiple platforms in a dynamic queue with visualization and data storage.
See @fig:fastapi-service.
{#fig:fastapi-service width=50%}
This software was developed end enhancing Cloudmesh, a suite of software to make using cloud and HPC resources easier. Specifically, we have added a library called Cloudmesh Controlled Computing (cloudmesh-cc) that adds workflow features to control the execution of jobs on remote compute resources.
The goal is to provide numerous methods of specifying the workflows on a local computer and running them on remote services such as HPC and cloud computing resources. This includes REST services and command line tools. The software developed is freely available and can easily be installed with standard Python tools so integration in the Python ecosystem using virtualenv's and Anaconda is simple.
A hybrid multi-cloud analytics service framework was created to manage heterogeneous and remote workflows, queues, and jobs. It was designed for access through both the command line and REST services to simplify the coordination of tasks on remote computers. In addition, this service supports multiple operating systems like macOS, Linux, and Windows 10 and 11, on various hosts: the computer's localhost, remote computers, and the Linux-based virtual image WSL. Jobs can be visualized and saved as a YAML and SVG data file. This workflow was extensively tested for functionality and reproducibility.
To test the workflow program, prepare a cm directory in your home directory by executing the following commands in a terminal:
mkdir ~/cm
cd ~/cm
pip install cloudmesh-installer -U
cloudmesh-installer get cc
cd cloudmesh-cc
pytest -v -x --capture=no tests/test_199_workflow_clean.py
This test runs three jobs within a singular workflow: the first job runs a local shell script, the second runs a local Python script, and the third runs a local Jupyter notebook.
The Modified National Institute of Standards and Technology Database is a machine learning database based on image processing Various MNIST files involving different machine learning cases were modified and tested on various local and remote machines These cases include Multilayer Perceptron, LSTM (Long short-term memory), Auto-Encoder, Convolutional, and Recurrent Neural Networks, Distributed Training, and PyTorch training.
See @fig:workflow-uml.
{#fig:workflow-uml}
The hybrid multi-cloud analytics service framework was created to ensure running jobs across many platforms. We designed a small and streamlined number of abstractions so that jobs and workflows can be represented easily. The design is flexible and can be expanded as each job can contain arbitrary arguments. This made it possible to custom design for each target type a specific job type so that execution on local and remote compute resources including batch operating systems can be achieved. The job types supported include: local job on Linux, macOS, Windows 10, and Windows 11, jobs running in WSL on Windows computers, remote jobs using ssh, and batch jobs using Slurm.
In addition, we leveraged the existing Networkx Graph framework to allow dependencies between jobs. This greatly reduced the complexity of the implementation while being able to leverage graphical displays of the workflow, as well as using scheduling jobs with for example topological sort available in Networkx. Custom schedulers can be designed easily based on the dependencies and job types managed through this straightforward interface. The status of the jobs is stored in a database that can be monitored during program execution. The creation of the jobs is done on the fly, e.g. when the job is needed to be determined on the dependencies when all its parents are resolved. This is especially important as it allows dynamic workflow patterns to be implemented while results from previous calculations can be used in later stages of the workflow.
We have developed a simple-to-use API for this so programs can be formulated using the API in Python. However, we embedded this API also in a prototype REST service to showcase that integration into language-independent frameworks is possible. The obvious functions to manage workflows are supported including graph specification through configuration files, upload of workflows, export, adding jobs and dependencies, and visualizing the workflow during the execution. An important feature that we added is the monitoring of the jobs while using progress reports through automated log file mining. This way each job reports the progress during the execution. This is especially of importance when we run very complex and long-running jobs.
The REST service was implemented in FastAPI to leverage a small but fast service that features a much smaller footprint for implementation and setup in contrast to other similar REST service frameworks using python.
This architectural component building this framework is depicted @fig:workflow-uml. The code is available in this repository and manual pages are provided on how to install it: cloudmesh-cc.
The main interaction with the workflow is through the command line. With the framework, researchers and scientists should be able to create jobs on their own, place them in the workflow, and run them on various types of computers.
In addition, developers and users can utilize the built-in OpenAPI graphical user interface to manage workflows between jobs. They can be uploaded as YAML files or individually added through the build-in debug framework.
Improvements to this project will include code cleanup and manual development.
A poster based on a pre-alpha version of this code is available as ppt and PDF file. However, that version is no longer valid and is superseded by much improved efforts. The code summarized in the pre-alpha version was mainly used to teach a number of students Python and how to work in a team
Please note also that the poster contains inaccurate statements and descriptions and should not be used as a reference to this work.
Continued work was in part funded by the NSF CyberTraining: CIC: CyberTraining for Students and Technologies from Generation Z with the award numbers 1829704 and 2200409. We like to thank the following contributors for their help and evaluation in a pre-alpha version of the code: Jackson Miskill, Alex Beck, Alison Lu. We are excited that this effort contributed significantly to their increased understanding of Python and how to develop in a team using the Python ecosystem.
Command cc
==========
::
Usage:
cc start [-c] [--reload] [--host=HOST] [--port=PORT]
cc stop
cc status
cc doc
cc test
cc workflow add [--name=NAME] [--job=JOB] ARGS...
cc workflow add [--name=NAME] --filename=FILENAME
cc workflow delete [--name=NAME] [--job=JOB]
cc workflow list [--name=NAME] [--job=JOB]
cc workflow run [--name=NAME] [--job=JOB] [--filename=FILENAME]
cc workflow [--name=NAME] --dependencies=DEPENDENCIES
cc workflow status --name=NAME [--output=OUTPUT]
cc workflow graph --name=NAME
cc workflow service add [--name=NAME] FILENAME
cc workflow service list [--name=NAME] [--job=JOB]
cc workflow service add [--name=NAME] [--job=JOB] ARGS...
cc workflow service run --name=NAME
This command does some useful things.
Arguments:
FILENAME a file name
JOB the name of a job that has been created
COMMAND the command that is associated with the job name
SCHEDULER designation of how jobs should be pulled from the queue
Options:
--reload reload for debugging
--host=HOST the host ip [default: 127.0.0.1]
--port=PORT the port [default: 8000]
-c flag to set host and port to 0.0.0.0:8000
--filename=FILENAME file that contains a workflow specification
--name=NAME the workflow name
--job=JOB the job name
--command=COMMAND a command to be added to a job
--scheduler=SCHEDULER the scheduling technique that is to be used
Description:
Please note that all arguments such as NAME and JOB can be comma
separated parameter expansions, so a command can be applied to multiple
workflows or jobs at the same time
NEW WORKFLOW CLI COMMANDS
Note that none of the CLI commands use the Workflow service. All actions
are performed directly in the command shell.
cc workflow add [--name=NAME] [--job=JOB] ARGS...
cc workflow delete [--name=NAME] --job=JOB
cc workflow list [--name=NAME] [--job=JOB]
cc workflow run [--name=NAME] [--job=JOB] [--filename=FILENAME]
cc workflow [--name=NAME] --dependencies=DEPENDENCIES
cc workflow status --name=NAME [--output=OUTPUT]
cc workflow graph --name=NAME
NEW WORKFLOW SERVICE COMMANDS
Note that for all service commands to function you need to have a running
server. In future, we need a mechanism how to set the hostname and port also
for the service commands. For the time being they are issues to
127.0.0.1:8000
SERVICE MANAGEMENT COMMANDS
cc start [--reload] [--host=HOST] [--port=PORT]
start the service. one can add the host and port so the service is
started with http://host:port. The default is 127.0.0.1:8000.
If -c is specified 0.0.0.0:8000 is used.
cc stop
stop the service
Currently this command is not supported.
cc status
returns the status of the service
cc doc
opens a web browser and displays the OpenAPI specification
cc test
issues a simple test to see if the service is running. This command
may be in future eliminated as the status should encompass such a test.
WORKFLOW MANAGEMENT COMMANDS
Each workflow can be identified by its name. Note that through
cms set workflow=NAME the default name of the workflow can be set.
If it is not set the default is `workflow`
cc workflow service add [--name=NAME] [--directory=DIR] FILENAME
adds a workflow with the given name from data included in the filename.
the underlying database will use that name and if not explicitly
specified the location of the data base will be
~/.cloudmesh/workflow/NAME/NAME.yaml
To identify the location a special configuration file will be placed in
~/.cloudmesh/workflow/config.yaml that contains the location of
the directories for the named workflows.
cc workflow service list [--name=NAME] [--job=JOB]
this command reacts dependent on which options we specify
If we do not specify anything the workflows will be listed.
If we specify a workflow name only that workflow will be listed
If we also specify a job the job will be listed.
If we only specif the job name, all jobs with that name from all
workflows will be returned. # this feature not implemented
cc workflow service add [--name=NAME] --job=JOB ARGS...
This command adds a job. with the specified arguments. A check
is returned and the user is alerted if arguments are missing
arguments are passe in ATTRIBUTE=VALUE fashion.
if the name of the workflow is committed the default workflow is used.
If no cob name is specified an automated number that is kept in the
config.yaml file will be used and the name will be job-n
cc workflow service delete [--name=NAME] --job=JOB
deletes the job in the specified workflow
cc workflow service run [--name=NAME]
runs the names workflow. If no name is provided the default
workflow is used.
THIS MAY BE OUTDATED
cc workflow NAME DEPENDENCIES
with workflow command you can add dependencies between jobs. The dependencies
are added to a named workflow. Multiple workflows can be added to create a
complex workflow.
The dependency specification is simply a comma separated list of job names
introducing a direct acyclic graph.
> cms cc workflow simple a,b,d
> cms cc workflow simple a,c,d
which will introduce a workflow
> a
> / \
> b c
> \ /
> d
cc workflow run NAME
runs the workflow with the given name
cc workflow graph NAME
creates a graph with the current status. Tasks that have been
executed will be augmented by metadata, such as runtime
cc workflow status NAME --output=OUTPUT
prints the status of the workflow in various formats including
table, json, yaml
> cms cc --parameter="a[1-2,5],a10"
> example on how to use Parameter.expand. See source code at
> https://github.com/cloudmesh/cloudmesh-cc/blob/main/cloudmesh/cc/command/cc.py
> prints the expanded parameter as a list
> ['a1', 'a2', 'a3', 'a4', 'a5', 'a10']
> cc exp --experiment=a=b,c=d
> example on how to use Parameter.arguments_to_dict. See source code at
> https://github.com/cloudmesh/cloudmesh-cc/blob/main/cloudmesh/cc/command/cc.py
> prints the parameter as dict
> {'a': 'b', 'c': 'd'}