Configure Experiments

How to configure experiments

Edit expdef_cxxx.conf, jobs_cxxx.conf and platforms_cxxx.conf in the conf folder of the experiment.

expdef_cxxx.conf contains:
  • Start dates, members and chunks (number and length).
  • Experiment project source: origin (version control system or path)
  • Project configuration file path.
jobs_cxxx.conf contains the workflow to be run:
  • Scripts to execute.
  • Dependencies between tasks.
  • Task requirements (processors, wallclock time…).
  • Platform to use.
platforms_cxxx.conf contains:
  • HPC, fat-nodes and supporting computers configuration.

Note

platforms_cxxx.conf is usually provided by technicians, users will only have to change login and accounting options for HPCs.

You may want to configure Autosubmit parameters for the experiment. Just edit autosubmit_cxxx.conf.

autosubmit_cxxx.conf contains:
  • Maximum number of jobs to be running at the same time at the HPC.
  • Time (seconds) between connections to the HPC queue scheduler to poll already submitted jobs status.
  • Number of retrials if a job fails.

Then, Autosubmit create command uses the expdef_cxxx.conf and generates the experiment: After editing the files you can proceed to the experiment workflow creation. Experiment workflow, which contains all the jobs and its dependencies, will be saved as a pkl file:

autosubmit create EXPID

EXPID is the experiment identifier.

Options:

usage: autosubmit create [-group_by {date,member,chunk,split} -expand -expand_status] [-h] [-np] [-cw] expid

  expid          experiment identifier

  -h, --help     show this help message and exit
  -np, --noplot  omit plot creation
  --hide,        hide the plot
  -group_by {date,member,chunk,split,automatic}
                        criteria to use for grouping jobs
  -expand,              list of dates/members/chunks to expand
  -expand_status,       status(es) to expand
  -nt                   --notransitive
                            prevents doing the transitive reduction when plotting the workflow
  -cw                   --check_wrapper
                            Generate the wrapper in the current workflow
  -d                    --detail
                            Shows Job List view in terminal

Example:

autosubmit create cxxx

In order to understand more the grouping options, which are used for visualization purposes, please check Grouping jobs.

More info on pickle can be found at http://docs.python.org/library/pickle.html

How to add a new job

To add a new job, open the <experiments_directory>/cxxx/conf/jobs_cxxx.conf file where cxxx is the experiment identifier and add this text:s

[new_job]
FILE = <new_job_template>

This will create a new job named “new_job” that will be executed once at the default platform. This job will user the template located at <new_job_template> (path is relative to project folder).

This is the minimum job definition and usually is not enough. You usually will need to add some others parameters:

  • PLATFORM: allows you to execute the job in a platform of your choice. It must be defined in the experiment’s platforms.conf file or to have the value ‘LOCAL’ that always refer to the machine running Autosubmit

  • RUNNING: defines if jobs runs only once or once per start-date, member or chunk. Options are: once, date, member, chunk

  • DEPENDENCIES: defines dependencies from job as a list of parents jobs separated by spaces. For example, if ‘new_job’ has to wait for “old_job” to finish, you must add the line “DEPENDENCIES = old_job”.

    • For dependencies to jobs running in previous chunks, members or start-dates, use -(DISTANCE). For example, for a job “SIM” waiting for
    the previous “SIM” job to finish, you have to add “DEPENDENCIES = SIM-1”.
    • For dependencies that are not mandatory for the normal workflow behaviour, you must add the char ‘?’ at the end of the dependency.

  • SELECT_CHUNKS (optional): by default, all sections depend on all jobs the items specified on the DEPENDENCIES parameter. However, with this parameter, you could select the chunks of a specific job section. At the end of this doc, you will find diverse examples of this feature. The syntax is as follows:

[jobs]
SELECT_CHUNKS = SIM*[1]*[3] # Enables the dependency of  chunk 1 with chunk 3. While chunks 2,4  won't be linked.
SELECT_CHUNKS = SIM*[1:3] # Enables the dependency of chunk 1,2 and 3. While 4 won't be linked.
SELECT_CHUNKS = SIM*[1,3] # Enables the dependency of chunk 1 and 3. While 2 and 4 won't be linked
SELECT_CHUNKS = SIM*[1] # Enables the dependency of chunk 1. While 2, 3 and 4 won't be linked
  • SELECT_MEMBERS (optional): by default, all sections depend on all jobs the items specified on the DEPENDENCIES parameter. However, with this parameter, you could select the members of a specific job section. At the end of this doc, you will find diverse examples of this feature. Caution, you must pick the member index, not the member name.
[expdef.conf]
...
MEMBERS = AA BB CC DD
...
[jobs.conf]
SELECT_MEMBERS = SIM*[1]*[3] # Enables the dependency of member BB with member DD. While AA and CC won't be linked.
SELECT_MEMBERS = SIM*[1:3] # Enables the dependency of member  BB,CC and DD. While AA won't be linked.
SELECT_MEMBERS = SIM*[1,3] # Enables the dependency of member BB and DD. While AA and CC won't be linked
SELECT_MEMBERS = SIM*[1] # Enables the dependency of member BB. While AA, CC and DD won't be linked
  • EXCLUDED_CHUNKS (optional): With this parameter, you can prevent the generation of jobs for a list of chunks.
  • EXCLUDED_MEMBERS (optional): With this parameter, you can prevent the generation of jobs for a list of members.

For jobs running in HPC platforms, usually you have to provide information about processors, wallclock times and more. To do this use:

  • WALLCLOCK: wallclock time to be submitted to the HPC queue in format HH:MM
  • PROCESSORS: processors number to be submitted to the HPC. If not specified, defaults to 1.
  • THREADS: threads number to be submitted to the HPC. If not specified, defaults to 1.
  • TASKS: tasks number to be submitted to the HPC. If not specified, defaults to 1.
  • NODES: nodes number to be submitted to the HPC. If not specified, the directive is not added.
  • HYPERTHREADING: Enables Hyper-threading, this will double the max amount of threads. defaults to false. ( Not available on slurm platforms )
  • QUEUE: queue to add the job to. If not specified, uses PLATFORM default.
  • RETRIALS: Number of retrials if job fails
  • DELAY_RETRY_TIME: Allows to put a delay between retries. Triggered when a job fails. If not specified, Autosubmit will retry the job as soon as possible. Accepted formats are: plain number (there will be a constant delay between retrials, of as many seconds as specified), plus (+) sign followed by a number (the delay will steadily increase by the addition of these number of seconds), or multiplication (*) sign follows by a number (the delay after n retries will be the number multiplied by 10*n). Having this in mind, the ideal scenario is to use +(number) or plain(number) in case that the HPC has little issues or the experiment will run for a little time. Otherwise, is better to use the *(number) approach.
#DELAY_RETRY_TIME = 11
#DELAY_RETRY_TIME = +11 # will wait 11 + number specified
#DELAY_RETRY_TIME = *11 # will wait 11,110,1110,11110...* by 10 to prevent a too big number

There are also other, less used features that you can use:

  • FREQUENCY: specifies that a job has only to be run after X dates, members or chunk. A job will always be created for the last one. If not specified, defaults to 1
  • SYNCHRONIZE: specifies that a job with RUNNING=chunk, has to synchronize its dependencies chunks at a ‘date’ or ‘member’ level, which means that the jobs will be unified: one per chunk for all members or dates. If not specified, the synchronization is for each chunk of all the experiment.
  • RERUN_ONLY: determines if a job is only to be executed in reruns. If not specified, defaults to false.
  • CUSTOM_DIRECTIVES: Custom directives for the HPC resource manager headers of the platform used for that job.
  • SKIPPABLE: When this is true, the job will be able to skip it work if there is an higher chunk or member already ready, running, queuing or in complete status.
  • EXPORT: Allows to run an env script or load some modules before running this job.
  • EXECUTABLE: Allows to wrap a job for be launched with a set of env variables.
  • QUEUE: queue to add the job to. If not specified, uses PLATFORM default.
  • EXTENDED_HEADER_PATH: path to a script to be appended at the begging of the .cmd script that Autosubmit generates. Only supports job type BASH.
  • EXTENDED_TAILER_PATH: path to a script to be appended at the end of the .cmd script that Autosubmit generates. Only supports job type BASH.

How to configure email notifications

To configure the email notifications, you have to follow two configuration steps:

  1. First you have to enable email notifications and set the accounts where you will receive it.

Edit autosubmit_cxxx.conf in the conf folder of the experiment.

Hint

Remember that you can define more than one email address divided by a whitespace.

Example:

vi <experiments_directory>/cxxx/conf/autosubmit_cxxx.conf

[mail]
# Enable mail notifications for remote_failures
# Default = True
NOTIFY_ON_REMOTE_FAIL = True
# Enable mail notifications
# Default = False
NOTIFICATIONS = True
# Mail address where notifications will be received
TO =  jsmith@example.com  rlewis@example.com
  1. Then you have to define for which jobs you want to be notified.

Edit jobs_cxxx.conf in the conf folder of the experiment.

Hint

You will be notified every time the job changes its status to one of the statuses defined on the parameter NOTIFY_ON

Hint

Remember that you can define more than one job status divided by a whitespace.

Example:

vi <experiments_directory>/cxxx/conf/jobs_cxxx.conf

[LOCAL_SETUP]
FILE = LOCAL_SETUP.sh
PLATFORM = LOCAL
NOTIFY_ON = FAILED COMPLETED

How to add a new platform

Hint

If you are interested in changing the communications library, go to the section below.

To add a new platform, open the <experiments_directory>/cxxx/conf/platforms_cxxx.conf file where cxxx is the experiment identifier and add this text:

[new_platform]
TYPE = <platform_type>
HOST = <host_name>
PROJECT = <project>
USER = <user>
SCRATCH = <scratch_dir>

This will create a platform named “new_platform”. The options specified are all mandatory:

  • TYPE: queue type for the platform. Options supported are PBS, SGE, PS, LSF, ecaccess and SLURM.
  • HOST: hostname of the platform
  • PROJECT: project for the machine scheduler
  • USER: user for the machine scheduler
  • SCRATCH_DIR: path to the scratch directory of the machine
  • VERSION: determines de version of the platform type

Warning

With some platform types, Autosubmit may also need the version, forcing you to add the parameter VERSION. These platforms are PBS (options: 10, 11, 12) and ecaccess (options: pbs, loadleveler).

Some platforms may require to run serial jobs in a different queue or platform. To avoid changing the job configuration, you can specify what platform or queue to use to run serial jobs assigned to this platform:

  • SERIAL_PLATFORM: if specified, Autosubmit will run jobs with only one processor in the specified platform.
  • SERIAL_QUEUE: if specified, Autosubmit will run jobs with only one processor in the specified queue. Autosubmit will ignore this configuration if SERIAL_PLATFORM is provided

There are some other parameters that you may need to specify:

  • BUDGET: budget account for the machine scheduler. If omitted, takes the value defined in PROJECT
  • ADD_PROJECT_TO_HOST = option to add project name to host. This is required for some HPCs
  • QUEUE: if given, Autosubmit will add jobs to the given queue instead of platform’s default queue
  • TEST_SUITE: if true, autosubmit test command can use this queue as a main queue. Defaults to false
  • MAX_WAITING_JOBS: maximum number of jobs to be waiting in this platform.
  • TOTAL_JOBS: maximum number of jobs to be running at the same time in this platform.
  • CUSTOM_DIRECTIVES: Custom directives for the resource manager of this platform.

Example:

[platform]
TYPE = SGE
HOST = hostname
PROJECT = my_project
ADD_PROJECT_TO_HOST = true
USER = my_user
SCRATCH_DIR = /scratch
TEST_SUITE = True
CUSTOM_DIRECTIVES = [ "my_directive" ]

How to request exclusivity or reservation

To request exclusivity or reservation for your jobs, you can configure two platform variables:

Edit platforms_cxxx.conf in the conf folder of the experiment.

Hint

Until now, it is only available for Marenostrum.

Hint

To define some jobs with exclusivity/reservation and some others without it, you can define twice a platform, one with this parameters and another one without it.

Example:

vi <experiments_directory>/cxxx/conf/platforms_cxxx.conf

[marenostrum3]
TYPE = LSF
HOST = mn-bsc32
PROJECT = bsc32
ADD_PROJECT_TO_HOST = false
USER = bsc32XXX
SCRATCH_DIR = /gpfs/scratch
TEST_SUITE = True
EXCLUSIVITY = True

Of course, you can configure only one or both. For example, for reservation it would be:

Example:

vi <experiments_directory>/cxxx/conf/platforms_cxxx.conf

[marenostrum3]
TYPE = LSF
...
RESERVATION = your-reservation-id

How to change the communications library

In order to handle the remote communications with the different platforms, Autosubmit uses an implementation of a communications library. There are multiple implementations, so you can choose any of them.

Hint

At this moment there are one available communication library which is paramiko.

To change the communications library, open the <experiments_directory>/cxxx/conf/autosubmit_cxxx.conf file where cxxx is the experiment identifier and change the value of the API configuration variable in the communications section:

[communications]
# Communications library used to connect with platforms: paramiko.
# Default = paramiko
API = paramiko

How to set a custom interpreter for your job

If the remote platform does not implement the interpreter you need, you can customize the shebang of your job script so it points to the relative path of the interpreter you want.

In the file:

vi <experiments_directory>/cxxx/conf/jobs_cxxx.conf

# Example job with all options specified

## Job name
# [JOBNAME]
## Script to execute. If not specified, job will be omitted from workflow.
## Path relative to the project directory
# FILE =
## Platform to execute the job. If not specified, defaults to HPCARCH in expedf file.
## LOCAL is always defined and refers to current machine
# PLATFORM =
## Queue to add the job to. If not specified, uses PLATFORM default.
# QUEUE =
## Defines dependencies from job as a list of parents jobs separated by spaces.
## Dependencies to jobs in previous chunk, member o startdate, use -(DISTANCE)
# DEPENDENCIES = INI SIM-1 CLEAN-2
## Define if jobs runs once, once per stardate, once per member or once per chunk. Options: once, date, member, chunk.
## If not specified, defaults to once
# RUNNING = once
## Specifies that job has only to be run after X dates, members or chunk. A job will always be created for the last
## If not specified, defaults to 1
# FREQUENCY = 3
## On a job with FREQUENCY > 1, if True, the dependencies are evaluated against all
## jobs in the frequency interval, otherwise only evaluate dependencies against current
## iteration.
## If not specified, defaults to True
# WAIT = False
## Defines if job is only to be executed in reruns. If not specified, defaults to false.
# RERUN_ONLY = False
## Wallclock to be submitted to the HPC queue in format HH:MM
# WALLCLOCK = 00:05
## Processors number to be submitted to the HPC. If not specified, defaults to 1.
## Wallclock chunk increase (WALLCLOCK will be increased according to the formula WALLCLOCK + WCHUNKINC * (chunk - 1)).
## Ideal for sequences of jobs that change their expected running time according to the current chunk.
# WCHUNKINC = 00:01
# PROCESSORS = 1
## Threads number to be submitted to the HPC. If not specified, defaults to 1.
# THREADS = 1
## Tasks number to be submitted to the HPC. If not specified, defaults to 1.
# Tasks = 1
## Enables hyper-threading. If not specified, defaults to false.
# HYPERTHREADING = false
## Memory requirements for the job in MB
# MEMORY = 4096
##  Number of retrials if a job fails. If not specified, defaults to the value given on experiment's autosubmit.conf
# RETRIALS = 4
##  Allows to put a delay between retries, of retrials if a job fails. If not specified, it will be static
# The ideal is to use the +(number) approach or plain(number) in case that the hpc platform has little issues or the experiment will run for a short period of time
# And *(10) in case that the filesystem is having large  delays or the experiment will run for a lot of time.
# DELAY_RETRY_TIME = 11
# DELAY_RETRY_TIME = +11 # will wait 11 + number specified
# DELAY_RETRY_TIME = *11 # will wait 11,110,1110,11110...* by 10 to prevent a too big number
## Some jobs can not be checked before running previous jobs. Set this option to false if that is the case
# CHECK = False
## Select the interpreter that will run the job. Options: bash, python, r Default: bash
# TYPE = bash
## Specify the path to the interpreter. If empty, use system default based on job type  . Default: empty
# EXECUTABLE = /my_python_env/python3

You can give a path to the EXECUTABLE setting of your job. Autosubmit will replace the shebang with the path you provided.

Example:

[POST]
FILE = POST.sh
DEPENDENCIES = SIM
RUNNING = chunk
WALLCLOCK = 00:05
EXECUTABLE = /my_python_env/python3

This job will use the python interpreter located in the relative path /my_python_env/python3/

It is also possible to use variables in the EXECUTABLE path.

Example:

[POST]
FILE = POST.sh
DEPENDENCIES = SIM
RUNNING = chunk
WALLCLOCK = 00:05
EXECUTABLE = %PROJDIR%/my_python_env/python3

The result is a shebang line #!/esarchive/autosubmit/my_python_env/python3.

How to create and run only selected members

Your experiment is defined and correctly configured, but you want to create it only considering some selected members, and also to avoid creating the whole experiment to run only the members you want. Then, you can do it by configuring the setting RUN_ONLY_MEMBERS in the file:

vi <experiments_directory>/cxxx/conf/expdef_cxxx.conf

[DEFAULT]
# Experiment identifier
# No need to change
EXPID = cxxx
# HPC name.
# No need to change
HPCARCH = ithaca

[experiment]
# Supply the list of start dates. Available formats: YYYYMMDD YYYYMMDDhh YYYYMMDDhhmm
# Also you can use an abbreviated syntax for multiple dates with common parts:
# 200001[01 15] <=> 20000101 20000115
# DATELIST = 19600101 19650101 19700101
# DATELIST = 1960[0101 0201 0301]
DATELIST = 19900101
# Supply the list of members. LIST = fc0 fc1 fc2 fc3 fc4
MEMBERS = fc0
# Chunk size unit. STRING = hour, day, month, year
CHUNKSIZEUNIT = month
# Chunk size. NUMERIC = 4, 6, 12
CHUNKSIZE = 1
# Total number of chunks in experiment. NUMERIC = 30, 15, 10
NUMCHUNKS = 2
# Calendar used. LIST: standard, noleap
CALENDAR = standard
# List of members that can be included in this run. Optional.
# RUN_ONLY_MEMBERS = fc0 fc1 fc2 fc3 fc4
# RUN_ONLY_MEMBERS = fc[0-4]
RUN_ONLY_MEMBERS =

You can set the RUN_ONLY_MEMBERS value as shown in the format examples above it. Then, Job List generation is performed as usual. However, an extra step is performed that will filter the jobs according to RUN_ONLY_MEMBERS. It discards jobs belonging to members not considered in the value provided, and also we discard these jobs from the dependency tree (parents and children). The filtered Job List is returned.

The necessary changes have been implemented in the API so you can correctly visualize experiments implementing this new setting in Autosubmit GUI.

Important

Wrappers are correctly formed considering the resulting jobs.

Remote Dependencies - Presubmission feature

There is also the possibility of setting the option PRESUBMISSION to True in the config directive. This allows more than one package containing simple or wrapped jobs to be submitted at the same time, even when the dependencies between jobs aren’t yet satisfied.

This is only useful for cases when the job scheduler considers the time a job has been queuing to determine the job’s priority (and the scheduler understands the dependencies set between the submitted packages). New packages can be created as long as the total number of jobs are below than the number defined in the TOTALJOBS variable.

The jobs that are waiting in the remote platform, will be marked as HOLD.

How to configure

In autosubmit_cxxx.conf, regardless of the how your workflow is configured.

For example:

[config]
EXPID = ....
AUTOSUBMIT_VERSION = 3.13.0b
...
PRESUBMISSION = TRUE
MAXWAITINGJOBS = 100
TOTALJOBS = 100
...