DPDispatcher’s documentation

DPDispatcher is a Python package used to generate HPC (High Performance Computing) scheduler systems (Slurm/PBS/LSF/dpcloudserver) jobs input scripts and submit these scripts to HPC systems and poke until they finish.

DPDispatcher will monitor (poke) until these jobs finish and download the results files (if these jobs is running on remote systems connected by SSH).

Install DPDispatcher

DPDispatcher can installed by pip:

pip install dpdispatcher

Getting Started

DPDispatcher provides the following classes:

• Task class, which represents a command to be run on batch job system, as well as the essential files need by the command.

• Submission class, which represents a collection of jobs defined by the HPC system. And there may be common files to be uploaded by them. DPDispatcher will create and submit these jobs when a submission instance execute run_submission method. This method will poke until the jobs finish and return.

• Job class, a class used by Submission class, which represents a job on the HPC system. Submission will generate jobs’ submitting scripts used by HPC systems automatically with the Task and Resources

• Resources class, which represents the computing resources for each job within a submission.

You can use DPDispatcher in a Python script to submit five tasks:

from dpdispatcher import Machine, Resources, Task, Submission

machine = Machine.load_from_json('machine.json')
resources = Resources.load_from_json('resources.json')

task0 = Task.load_from_json('task.json')

task1 = Task(command='cat example.txt', task_work_path='dir1/', forward_files=['example.txt'], backward_files=['out.txt'], outlog='out.txt')
task2 = Task(command='cat example.txt', task_work_path='dir2/', forward_files=['example.txt'], backward_files=['out.txt'], outlog='out.txt')
task3 = Task(command='cat example.txt', task_work_path='dir3/', forward_files=['example.txt'], backward_files=['out.txt'], outlog='out.txt')
task4 = Task(command='cat example.txt', task_work_path='dir4/', forward_files=['example.txt'], backward_files=['out.txt'], outlog='out.txt')

task_list = [task0, task1, task2, task3, task4]

submission = Submission(work_base='lammps_md_300K_5GPa/',
    machine=machine, 
    resources=resources,
    task_list=task_list,
    forward_common_files=['graph.pb'], 
    backward_common_files=[]
)

submission.run_submission()

where machine.json is

{
    "batch_type": "Slurm",
    "context_type": "SSHContext",
    "local_root" : "/home/user123/workplace/22_new_project/",
    "remote_root": "/home/user123/dpdispatcher_work_dir/",
    "remote_profile":{
        "hostname": "39.106.xx.xxx",
        "username": "user123",
        "port": 22,
        "timeout": 10
    }
}

resources.json is

{
    "number_node": 1,
    "cpu_per_node": 4,
    "gpu_per_node": 1,
    "queue_name": "GPUV100",
    "group_size": 5
}

and task.json is

{
    "command": "lmp -i input.lammps",
    "task_work_path": "bct-0/",
    "forward_files": [
        "conf.lmp",
        "input.lammps"
    ],
    "backward_files": [
        "log.lammps"
    ],
    "outlog": "log",
    "errlog": "err",
}

You may also submit mutiple GPU jobs: complex resources example

resources = Resources(
    number_node=1,
    cpu_per_node=4,
    gpu_per_node=2,
    queue_name="GPU_2080Ti",
    group_size=4,
    custom_flags=[
        "#SBATCH --nice=100", 
        "#SBATCH --time=24:00:00"
    ],
    strategy={
        # used when you want to add CUDA_VISIBLE_DIVECES automatically
        "if_cuda_multi_devices": True 
    },
    para_deg=1,
    # will unload these modules before running tasks
    module_unload_list=["singularity"],
    # will load these modules before running tasks
    module_list=["singularity/3.0.0"],
    # will source the environment files before running tasks
    source_list=["./slurm_test.env"],
    # the envs option is used to export environment variables
    # And it will generate a line like below.
    # export DP_DISPATCHER_EXPORT=test_foo_bar_baz
    envs={"DP_DISPATCHER_EXPORT": "test_foo_bar_baz"},
)

The details of parameters can be found in Machine Parameters, Resources Parameters, and Task Parameters.

Machine parameters

Note

One can load, modify, and export the input file by using our effective web-based tool DP-GUI. All parameters below can be set in DP-GUI. By clicking “SAVE JSON”, one can download the input file.

machine:

type: dict

argument path: machine

batch_type:

type: str

argument path: machine/batch_type

The batch job system type. Option: DpCloudServer, Shell, SlurmJobArray, Torque, DistributedShell, Lebesgue, PBS, LSF, Slurm

local_root:

type: str

argument path: machine/local_root

The dir where the tasks and relating files locate. Typically the project dir.

remote_root:

type: str, optional

argument path: machine/remote_root

The dir where the tasks are executed on the remote machine. Only needed when context is not lazy-local.

clean_asynchronously:

type: bool, optional, default: False

argument path: machine/clean_asynchronously

Clean the remote directory asynchronously after the job finishes.

Depending on the value of context_type, different sub args are accepted.

context_type:

type: str (flag key)

argument path: machine/context_type

possible choices: SSHContext, LazyLocalContext, LocalContext, DpCloudServerContext, HDFSContext, LebesgueContext

The connection used to remote machine. Option: SSHContext, HDFSContext, LocalContext, DpCloudServerContext, LazyLocalContext, LebesgueContext

When context_type is set to SSHContext (or its aliases sshcontext, SSH, ssh):

remote_profile:

type: dict

argument path: machine[SSHContext]/remote_profile

The information used to maintain the connection with remote machine.

hostname:

type: str

argument path: machine[SSHContext]/remote_profile/hostname

hostname or ip of ssh connection.

username:

type: str

argument path: machine[SSHContext]/remote_profile/username

username of target linux system

password:

type: str, optional

argument path: machine[SSHContext]/remote_profile/password

(deprecated) password of linux system. Please use SSH keys instead to improve security.

port:

type: int, optional, default: 22

argument path: machine[SSHContext]/remote_profile/port

ssh connection port.

key_filename:

type: NoneType | str, optional, default: None

argument path: machine[SSHContext]/remote_profile/key_filename

key filename used by ssh connection. If left None, find key in ~/.ssh or use password for login

passphrase:

type: NoneType | str, optional, default: None

argument path: machine[SSHContext]/remote_profile/passphrase

passphrase of key used by ssh connection

timeout:

type: int, optional, default: 10

argument path: machine[SSHContext]/remote_profile/timeout

timeout of ssh connection

totp_secret:

type: NoneType | str, optional, default: None

argument path: machine[SSHContext]/remote_profile/totp_secret

Time-based one time password secret. It should be a base32-encoded string extracted from the 2D code.

When context_type is set to LazyLocalContext (or its aliases lazylocalcontext, LazyLocal, lazylocal):

remote_profile:

type: dict, optional

argument path: machine[LazyLocalContext]/remote_profile

The information used to maintain the connection with remote machine. This field is empty for this context.

When context_type is set to LocalContext (or its aliases localcontext, Local, local):

remote_profile:

type: dict, optional

argument path: machine[LocalContext]/remote_profile

The information used to maintain the connection with remote machine. This field is empty for this context.

When context_type is set to DpCloudServerContext (or its aliases dpcloudservercontext, DpCloudServer, dpcloudserver):

remote_profile:

type: dict

argument path: machine[DpCloudServerContext]/remote_profile

The information used to maintain the connection with remote machine.

email:

type: str

argument path: machine[DpCloudServerContext]/remote_profile/email

Email

password:

type: str

argument path: machine[DpCloudServerContext]/remote_profile/password

Password

program_id:

type: int

argument path: machine[DpCloudServerContext]/remote_profile/program_id

Program ID

input_data:

type: dict

argument path: machine[DpCloudServerContext]/remote_profile/input_data

Configuration of job

When context_type is set to HDFSContext (or its aliases hdfscontext, HDFS, hdfs):

remote_profile:

type: dict, optional

argument path: machine[HDFSContext]/remote_profile

The information used to maintain the connection with remote machine. This field is empty for this context.

When context_type is set to LebesgueContext (or its aliases lebesguecontext, Lebesgue, lebesgue):

remote_profile:

type: dict

argument path: machine[LebesgueContext]/remote_profile

The information used to maintain the connection with remote machine.

email:

type: str

argument path: machine[LebesgueContext]/remote_profile/email

Email

password:

type: str

argument path: machine[LebesgueContext]/remote_profile/password

Password

program_id:

type: int

argument path: machine[LebesgueContext]/remote_profile/program_id

Program ID

input_data:

type: dict

argument path: machine[LebesgueContext]/remote_profile/input_data

Configuration of job

Resources parameters

Note

One can load, modify, and export the input file by using our effective web-based tool DP-GUI. All parameters below can be set in DP-GUI. By clicking “SAVE JSON”, one can download the input file for.

resources:

type: dict

argument path: resources

number_node:

type: int

argument path: resources/number_node

The number of node need for each job

cpu_per_node:

type: int

argument path: resources/cpu_per_node

cpu numbers of each node assigned to each job.

gpu_per_node:

type: int

argument path: resources/gpu_per_node

gpu numbers of each node assigned to each job.

queue_name:

type: str

argument path: resources/queue_name

The queue name of batch job scheduler system.

group_size:

type: int

argument path: resources/group_size

The number of tasks in a job. 0 means infinity.

custom_flags:

type: list, optional

argument path: resources/custom_flags

The extra lines pass to job submitting script header

strategy:

type: dict, optional

argument path: resources/strategy

strategies we use to generation job submitting scripts.

if_cuda_multi_devices:

type: bool, optional, default: False

argument path: resources/strategy/if_cuda_multi_devices

ratio_unfinished:

type: float, optional, default: 0.0

argument path: resources/strategy/ratio_unfinished

para_deg:

type: int, optional, default: 1

argument path: resources/para_deg

Decide how many tasks will be run in parallel.

source_list:

type: list, optional, default: []

argument path: resources/source_list

The env file to be sourced before the command execution.

module_purge:

type: bool, optional, default: False

argument path: resources/module_purge

Remove all modules on HPC system before module load (module_list)

module_unload_list:

type: list, optional, default: []

argument path: resources/module_unload_list

The modules to be unloaded on HPC system before submitting jobs

module_list:

type: list, optional, default: []

argument path: resources/module_list

The modules to be loaded on HPC system before submitting jobs

envs:

type: dict, optional, default: {}

argument path: resources/envs

The environment variables to be exported on before submitting jobs

wait_time:

type: int | float, optional, default: 0

argument path: resources/wait_time

The waitting time in second after a single task submitted

Depending on the value of batch_type, different sub args are accepted.

batch_type:

type: str (flag key)

argument path: resources/batch_type

possible choices: SlurmJobArray, PBS, Slurm, Torque, DpCloudServer, Lebesgue, DistributedShell, Shell, LSF

The batch job system type loaded from machine/batch_type.

When batch_type is set to SlurmJobArray (or its alias slurmjobarray):

kwargs:

type: dict, optional

argument path: resources[SlurmJobArray]/kwargs

Extra arguments.

custom_gpu_line:

type: NoneType | str, optional, default: None

argument path: resources[SlurmJobArray]/kwargs/custom_gpu_line

Custom GPU configuration, starting with #SBATCH

When batch_type is set to PBS (or its alias pbs):

kwargs:

type: dict, optional

argument path: resources[PBS]/kwargs

This field is empty for this batch.

When batch_type is set to Slurm (or its alias slurm):

kwargs:

type: dict, optional

argument path: resources[Slurm]/kwargs

Extra arguments.

custom_gpu_line:

type: NoneType | str, optional, default: None

argument path: resources[Slurm]/kwargs/custom_gpu_line

Custom GPU configuration, starting with #SBATCH

When batch_type is set to Torque (or its alias torque):

kwargs:

type: dict, optional

argument path: resources[Torque]/kwargs

This field is empty for this batch.

When batch_type is set to DpCloudServer (or its alias dpcloudserver):

kwargs:

type: dict, optional

argument path: resources[DpCloudServer]/kwargs

This field is empty for this batch.

When batch_type is set to Lebesgue (or its alias lebesgue):

kwargs:

type: dict, optional

argument path: resources[Lebesgue]/kwargs

This field is empty for this batch.

When batch_type is set to DistributedShell (or its alias distributedshell):

kwargs:

type: dict, optional

argument path: resources[DistributedShell]/kwargs

This field is empty for this batch.

When batch_type is set to Shell (or its alias shell):

kwargs:

type: dict, optional

argument path: resources[Shell]/kwargs

This field is empty for this batch.

When batch_type is set to LSF (or its alias lsf):

kwargs:

type: dict

argument path: resources[LSF]/kwargs

Extra arguments.

gpu_usage:

type: bool, optional, default: False

argument path: resources[LSF]/kwargs/gpu_usage

Choosing if GPU is used in the calculation step.

gpu_new_syntax:

type: bool, optional, default: False

argument path: resources[LSF]/kwargs/gpu_new_syntax

For LFS >= 10.1.0.3, new option -gpu for #BSUB could be used. If False, and old syntax would be used.

gpu_exclusive:

type: bool, optional, default: True

argument path: resources[LSF]/kwargs/gpu_exclusive

Only take effect when new syntax enabled. Control whether submit tasks in exclusive way for GPU.

custom_gpu_line:

type: NoneType | str, optional, default: None

argument path: resources[LSF]/kwargs/custom_gpu_line

Custom GPU configuration, starting with #BSUB

Task parameters

Note

One can load, modify, and export the input file by using our effective web-based tool DP-GUI. All parameters below can be set in DP-GUI. By clicking “SAVE JSON”, one can download the input file.

task:

type: dict

argument path: task

command:

type: str

argument path: task/command

A command to be executed of this task. The expected return code is 0.

task_work_path:

type: str

argument path: task/task_work_path

The dir where the command to be executed.

forward_files:

type: list

argument path: task/forward_files

The files to be uploaded in task_work_path before the task exectued.

backward_files:

type: list

argument path: task/backward_files

The files to be download to local_root in task_work_path after the task finished

outlog:

type: NoneType | str

argument path: task/outlog

The out log file name. redirect from stdout

errlog:

type: NoneType | str

argument path: task/errlog

The err log file name. redirect from stderr

DPDispatcher API

dpdispatcher package

dpdispatcher.info()

Subpackages

dpdispatcher.dpcloudserver package

Submodules

dpdispatcher.dpcloudserver.api module

class dpdispatcher.dpcloudserver.api.API(email, password)

Bases: object

Methods

check_job_has_uploaded
download
download_from_url
get
get_job_result_url
get_jobs
get_tasks
get_tasks_list
job_create
post
refresh_token
upload

check_job_has_uploaded(job_id)

download(oss_file, save_file, endpoint, bucket_name)

download_from_url(url, save_file)

get(url, params, retry=0)

get_job_result_url(job_id)

get_jobs(page=1, per_page=10)

get_tasks(job_id, group_id, page=1, per_page=10)

get_tasks_list(group_id, per_page=30)

job_create(job_type, oss_path, input_data, program_id=None, group_id=None)

post(url, params, retry=0)

refresh_token()

upload(oss_task_zip, zip_task_file, endpoint, bucket_name)

dpdispatcher.dpcloudserver.config module

dpdispatcher.dpcloudserver.retcode module

class dpdispatcher.dpcloudserver.retcode.RETCODE

Bases: object

DATAERR = '2002'

DBERR = '2000'

IOERR = '2003'

NODATA = '2300'

OK = '0000'

PARAMERR = '2101'

PWDERR = '2104'

REQERR = '2200'

ROLEERR = '2103'

THIRDERR = '2001'

TOKENINVALID = '2100'

UNDERDEBUG = '2301'

UNKOWNERR = '2400'

USERERR = '2102'

VERIFYERR = '2105'

dpdispatcher.dpcloudserver.temp_test module

dpdispatcher.dpcloudserver.zip_file module

dpdispatcher.dpcloudserver.zip_file.unzip_file(zip_file, out_dir='./')

dpdispatcher.dpcloudserver.zip_file.zip_file_list(root_path, zip_filename, file_list=[])

Submodules

dpdispatcher.JobStatus module

class dpdispatcher.JobStatus.JobStatus(value)

Bases: IntEnum

An enumeration.

completing = 6

finished = 5

running = 3

terminated = 4

unknown = 100

unsubmitted = 1

waiting = 2

dpdispatcher.arginfo module

dpdispatcher.base_context module

class dpdispatcher.base_context.BaseContext(*args, **kwargs)

Bases: object

Methods

machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
check_finish
clean
download
kill
load_from_dict
read_file
upload
write_file

bind_submission(submission)

check_finish(proc)

abstract clean()

abstract download(submission, check_exists=False, mark_failure=True, back_error=False)

kill(proc)

classmethod load_from_dict(context_dict)

classmethod machine_arginfo() → Argument

Generate the machine arginfo.

Returns

Argument

machine arginfo

classmethod machine_subfields() → List[Argument]

Generate the machine subfields.

Returns

list[Argument]

machine subfields

options = {'DpCloudServerContext', 'HDFSContext', 'LazyLocalContext', 'LebesgueContext', 'LocalContext', 'SSHContext'}

abstract read_file(fname)

subclasses_dict = {'DpCloudServer': <class 'dpdispatcher.dp_cloud_server_context.DpCloudServerContext'>, 'DpCloudServerContext': <class 'dpdispatcher.dp_cloud_server_context.DpCloudServerContext'>, 'HDFS': <class 'dpdispatcher.hdfs_context.HDFSContext'>, 'HDFSContext': <class 'dpdispatcher.hdfs_context.HDFSContext'>, 'LazyLocal': <class 'dpdispatcher.lazy_local_context.LazyLocalContext'>, 'LazyLocalContext': <class 'dpdispatcher.lazy_local_context.LazyLocalContext'>, 'Lebesgue': <class 'dpdispatcher.dp_cloud_server_context.LebesgueContext'>, 'LebesgueContext': <class 'dpdispatcher.dp_cloud_server_context.LebesgueContext'>, 'Local': <class 'dpdispatcher.local_context.LocalContext'>, 'LocalContext': <class 'dpdispatcher.local_context.LocalContext'>, 'SSH': <class 'dpdispatcher.ssh_context.SSHContext'>, 'SSHContext': <class 'dpdispatcher.ssh_context.SSHContext'>, 'dpcloudserver': <class 'dpdispatcher.dp_cloud_server_context.DpCloudServerContext'>, 'dpcloudservercontext': <class 'dpdispatcher.dp_cloud_server_context.DpCloudServerContext'>, 'hdfs': <class 'dpdispatcher.hdfs_context.HDFSContext'>, 'hdfscontext': <class 'dpdispatcher.hdfs_context.HDFSContext'>, 'lazylocal': <class 'dpdispatcher.lazy_local_context.LazyLocalContext'>, 'lazylocalcontext': <class 'dpdispatcher.lazy_local_context.LazyLocalContext'>, 'lebesgue': <class 'dpdispatcher.dp_cloud_server_context.LebesgueContext'>, 'lebesguecontext': <class 'dpdispatcher.dp_cloud_server_context.LebesgueContext'>, 'local': <class 'dpdispatcher.local_context.LocalContext'>, 'localcontext': <class 'dpdispatcher.local_context.LocalContext'>, 'ssh': <class 'dpdispatcher.ssh_context.SSHContext'>, 'sshcontext': <class 'dpdispatcher.ssh_context.SSHContext'>}

abstract upload(submission)

abstract write_file(fname, write_str)

dpdispatcher.distributed_shell module

class dpdispatcher.distributed_shell.DistributedShell(*args, **kwargs)

Bases: Machine

Methods

do_submit(job)	submit th job to yarn using distributed shell
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job)

do_submit(job)

submit th job to yarn using distributed shell

Parameters

jobJob class instance

job to be submitted

Returns

job_id: string

submit process id

gen_script_end(job)

gen_script_env(job)

gen_script_header(job)

dpdispatcher.dp_cloud_server module

class dpdispatcher.dp_cloud_server.DpCloudServer(*args, **kwargs)

Bases: Machine

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_local_script
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
map_dp_job_state
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_if_recover(submission)

check_status(job)

do_submit(job)

submit a single job, assuming that no job is running there.

gen_local_script(job)

gen_script(job)

gen_script_header(job)

static map_dp_job_state(status)

class dpdispatcher.dp_cloud_server.Lebesgue(*args, **kwargs)

Bases: DpCloudServer

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_local_script
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
map_dp_job_state
serialize
sub_script_cmd
sub_script_head

dpdispatcher.dp_cloud_server_context module

class dpdispatcher.dp_cloud_server_context.DpCloudServerContext(*args, **kwargs)

Bases: BaseContext

Methods

machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
check_file_exists
check_finish
check_home_file_exits
clean
download
kill
load_from_dict
read_file
read_home_file
upload
write_file
write_home_file
write_local_file

bind_submission(submission)

check_file_exists(fname)

check_home_file_exits(fname)

clean()

download(submission)

kill(cmd_pipes)

classmethod load_from_dict(context_dict)

classmethod machine_subfields() → List[Argument]

Generate the machine subfields.

Returns

list[Argument]

machine subfields

read_file(fname)

read_home_file(fname)

upload(submission)

write_file(fname, write_str)

write_home_file(fname, write_str)

write_local_file(fname, write_str)

class dpdispatcher.dp_cloud_server_context.LebesgueContext(*args, **kwargs)

Bases: DpCloudServerContext

Methods

machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
check_file_exists
check_finish
check_home_file_exits
clean
download
kill
load_from_dict
read_file
read_home_file
upload
write_file
write_home_file
write_local_file

dpdispatcher.dpdisp module

dpdispatcher.dpdisp.main()

dpdispatcher.hdfs_cli module

class dpdispatcher.hdfs_cli.HDFS

Bases: object

Fundamental class for HDFS basic manipulation

Methods

copy_from_local(local_path, to_uri)	Returns: True on success Raises: on unexpected error
exists(uri)	Check existence of hdfs uri Returns: True on exists Raises: RuntimeError
mkdir(uri)	Make new hdfs directory Returns: True on success Raises: RuntimeError
remove(uri)	Check existence of hdfs uri Returns: True on exists Raises: RuntimeError

copy_to_local
move
read_hdfs_file

static copy_from_local(local_path, to_uri)

Returns: True on success Raises: on unexpected error

static copy_to_local(from_uri, local_path)

static exists(uri)

Check existence of hdfs uri Returns: True on exists Raises: RuntimeError

static mkdir(uri)

Make new hdfs directory Returns: True on success Raises: RuntimeError

static move(from_uri, to_uri)

static read_hdfs_file(uri)

static remove(uri)

Check existence of hdfs uri Returns: True on exists Raises: RuntimeError

dpdispatcher.hdfs_context module

class dpdispatcher.hdfs_context.HDFSContext(*args, **kwargs)

Bases: BaseContext

Methods

download(submission[, check_exists, ...])	download backward files from HDFS root dir
machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.
upload(submission[, dereference])	upload forward files and forward command files to HDFS root dir

bind_submission
check_file_exists
check_finish
clean
get_job_root
kill
load_from_dict
read_file
write_file

bind_submission(submission)

check_file_exists(fname)

clean()

download(submission, check_exists=False, mark_failure=True, back_error=False)

download backward files from HDFS root dir

Parameters

submissionSubmission class instance

represents a collection of tasks, such as backward file names

Returns

none

get_job_root()

kill(job_id)

classmethod load_from_dict(context_dict)

read_file(fname)

upload(submission, dereference=True)

upload forward files and forward command files to HDFS root dir

Parameters

submissionSubmission class instance

represents a collection of tasks, such as forward file names

Returns

none

write_file(fname, write_str)

dpdispatcher.lazy_local_context module

class dpdispatcher.lazy_local_context.LazyLocalContext(*args, **kwargs)

Bases: BaseContext

Methods

machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
block_call
block_checkcall
call
check_file_exists
check_finish
clean
download
get_job_root
get_return
kill
load_from_dict
read_file
upload
write_file

bind_submission(submission)

block_call(cmd)

block_checkcall(cmd)

call(cmd)

check_file_exists(fname)

check_finish(proc)

clean()

download(jobs, check_exists=False, mark_failure=True, back_error=False)

get_job_root()

get_return(proc)

kill(job_id)

classmethod load_from_dict(context_dict)

read_file(fname)

upload(jobs, dereference=True)

write_file(fname, write_str)

class dpdispatcher.lazy_local_context.SPRetObj(ret)

Bases: object

Methods

read
readlines

read()

readlines()

dpdispatcher.local_context module

class dpdispatcher.local_context.LocalContext(*args, **kwargs)

Bases: BaseContext

Methods

machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
block_call
block_checkcall
call
check_file_exists
check_finish
clean
download
download_
get_job_root
get_return
kill
load_from_dict
read_file
upload
upload_
write_file

bind_submission(submission)

block_call(cmd)

block_checkcall(cmd)

call(cmd)

check_file_exists(fname)

check_finish(proc)

clean()

download(submission, check_exists=False, mark_failure=True, back_error=False)

download_(job_dirs, remote_down_files, check_exists=False, mark_failure=True, back_error=False)

get_job_root()

get_return(proc)

kill(job_id)

classmethod load_from_dict(context_dict)

read_file(fname)

upload(submission)

upload_(job_dirs, local_up_files, dereference=True)

write_file(fname, write_str)

class dpdispatcher.local_context.SPRetObj(ret)

Bases: object

Methods

read
readlines

read()

readlines()

dpdispatcher.lsf module

class dpdispatcher.lsf.LSF(*args, **kwargs)

Bases: Machine

LSF batch

Methods

default_resources(resources)
do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job)

default_resources(resources)

do_submit(job)

submit a single job, assuming that no job is running there.

gen_script(job)

gen_script_header(job)

classmethod resources_subfields() → List[Argument]

Generate the resources subfields.

Returns

list[Argument]

resources subfields

sub_script_cmd(res)

sub_script_head(res)

dpdispatcher.machine module

class dpdispatcher.machine.Machine(*args, **kwargs)

Bases: object

A machine is used to handle the connection with remote machines.

Parameters

contextSubClass derived from BaseContext

The context is used to mainatin the connection with remote machine.

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

classmethod arginfo()

bind_context(context)

abstract check_finish_tag(**kwargs)

check_if_recover(submission)

abstract check_status(job)

default_resources(res)

classmethod deserialize(machine_dict)

abstract do_submit(job)

submit a single job, assuming that no job is running there.

gen_command_env_cuda_devices(resources)

gen_script(job)

gen_script_command(job)

gen_script_custom_flags_lines(job)

gen_script_end(job)

gen_script_env(job)

abstract gen_script_header(job)

gen_script_wait(resources)

classmethod load_from_dict(machine_dict)

classmethod load_from_json(json_path)

options = {'DistributedShell', 'DpCloudServer', 'LSF', 'Lebesgue', 'PBS', 'Shell', 'Slurm', 'SlurmJobArray', 'Torque'}

classmethod resources_arginfo() → Argument

Generate the resources arginfo.

Returns

Argument

resources arginfo

classmethod resources_subfields() → List[Argument]

Generate the resources subfields.

Returns

list[Argument]

resources subfields

serialize(if_empty_remote_profile=False)

sub_script_cmd(res)

sub_script_head(res)

subclasses_dict = {'DistributedShell': <class 'dpdispatcher.distributed_shell.DistributedShell'>, 'DpCloudServer': <class 'dpdispatcher.dp_cloud_server.DpCloudServer'>, 'LSF': <class 'dpdispatcher.lsf.LSF'>, 'Lebesgue': <class 'dpdispatcher.dp_cloud_server.Lebesgue'>, 'PBS': <class 'dpdispatcher.pbs.PBS'>, 'Shell': <class 'dpdispatcher.shell.Shell'>, 'Slurm': <class 'dpdispatcher.slurm.Slurm'>, 'SlurmJobArray': <class 'dpdispatcher.slurm.SlurmJobArray'>, 'Torque': <class 'dpdispatcher.pbs.Torque'>, 'distributedshell': <class 'dpdispatcher.distributed_shell.DistributedShell'>, 'dpcloudserver': <class 'dpdispatcher.dp_cloud_server.DpCloudServer'>, 'lebesgue': <class 'dpdispatcher.dp_cloud_server.Lebesgue'>, 'lsf': <class 'dpdispatcher.lsf.LSF'>, 'pbs': <class 'dpdispatcher.pbs.PBS'>, 'shell': <class 'dpdispatcher.shell.Shell'>, 'slurm': <class 'dpdispatcher.slurm.Slurm'>, 'slurmjobarray': <class 'dpdispatcher.slurm.SlurmJobArray'>, 'torque': <class 'dpdispatcher.pbs.Torque'>}

dpdispatcher.pbs module

class dpdispatcher.pbs.PBS(*args, **kwargs)

Bases: Machine

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job)

default_resources(resources)

do_submit(job)

submit a single job, assuming that no job is running there.

gen_script(job)

gen_script_header(job)

class dpdispatcher.pbs.Torque(*args, **kwargs)

Bases: PBS

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_status(job)

dpdispatcher.shell module

class dpdispatcher.shell.Shell(*args, **kwargs)

Bases: Machine

Methods

do_submit(job)	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job)

default_resources(resources)

do_submit(job)

submit a single job, assuming that no job is running there.

gen_script(job)

gen_script_header(job)

dpdispatcher.slurm module

class dpdispatcher.slurm.Slurm(*args, **kwargs)

Bases: Machine

Methods

do_submit(job[, retry, max_retry])	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job, retry=0, max_retry=3)

default_resources(resources)

do_submit(job, retry=0, max_retry=3)

submit a single job, assuming that no job is running there.

gen_script(job)

gen_script_header(job)

classmethod resources_subfields() → List[Argument]

Generate the resources subfields.

Returns

list[Argument]

resources subfields

class dpdispatcher.slurm.SlurmJobArray(*args, **kwargs)

Bases: Slurm

Slurm with job array enabled for multiple tasks in a job

Methods

do_submit(job[, retry, max_retry])	submit a single job, assuming that no job is running there.
resources_arginfo()	Generate the resources arginfo.
resources_subfields()	Generate the resources subfields.

arginfo
bind_context
check_finish_tag
check_if_recover
check_status
default_resources
deserialize
gen_command_env_cuda_devices
gen_script
gen_script_command
gen_script_custom_flags_lines
gen_script_end
gen_script_env
gen_script_header
gen_script_wait
load_from_dict
load_from_json
serialize
sub_script_cmd
sub_script_head

check_finish_tag(job)

check_status(job, retry=0, max_retry=3)

gen_script_command(job)

gen_script_end(job)

gen_script_header(job)

dpdispatcher.ssh_context module

class dpdispatcher.ssh_context.SSHContext(*args, **kwargs)

Bases: BaseContext

Attributes

sftp

ssh

Methods

block_checkcall(cmd[, asynchronously, ...])	Run command with arguments.
machine_arginfo()	Generate the machine arginfo.
machine_subfields()	Generate the machine subfields.

bind_submission
block_call
call
check_file_exists
check_finish
clean
close
download
get_job_root
get_return
kill
load_from_dict
read_file
upload
write_file

bind_submission(submission)

block_call(cmd)

block_checkcall(cmd, asynchronously=False, stderr_whitelist=None)

Run command with arguments. Wait for command to complete. If the return code was zero then return, otherwise raise RuntimeError.

Parameters

cmd: str

The command to run.

asynchronously: bool, optional, default=False

Run command asynchronously. If True, nohup will be used to run the command.

call(cmd)

check_file_exists(fname)

check_finish(cmd_pipes)

clean()

close()

download(submission, check_exists=False, mark_failure=True, back_error=False)

get_job_root()

get_return(cmd_pipes)

kill(cmd_pipes)

classmethod load_from_dict(context_dict)

classmethod machine_subfields() → List[Argument]

Generate the machine subfields.

Returns

list[Argument]

machine subfields

read_file(fname)

property sftp

property ssh

upload(submission, dereference=True)

write_file(fname, write_str)

class dpdispatcher.ssh_context.SSHSession(hostname, username, password=None, port=22, key_filename=None, passphrase=None, timeout=10, totp_secret=None)

Bases: object

Attributes

remote

rsync_available

sftp

Returns sftp.

Methods

exec_command(cmd[, retry])

Calling self.ssh.exec_command but has an exception check.

arginfo
close
ensure_alive
get
get_ssh_client
put

static arginfo()

close()

ensure_alive(max_check=10, sleep_time=10)

exec_command(cmd, retry=0)

Calling self.ssh.exec_command but has an exception check.

get(from_f, to_f)

get_ssh_client()

put(from_f, to_f)

property remote: str

property rsync_available: bool

property sftp

Returns sftp. Open a new one if not existing.

dpdispatcher.submission module

class dpdispatcher.submission.Job(job_task_list, *, resources, machine=None)

Bases: object

Job is generated by Submission automatically. A job ususally has many tasks and it may request computing resources from job scheduler systems. Each Job can generate a script file to be submitted to the job scheduler system or executed locally.

Parameters

job_task_listlist of Task

the tasks belonging to the job

resourcesResources

the machine resources. Passed from Submission when it constructs jobs.

machinemachine

machine object to execute the job. Passed from Submission when it constructs jobs.

Methods

deserialize(job_dict[, machine])	convert the job_dict to a Submission class object
get_job_state()	get the jobs.
serialize([if_static])	convert the Task class instance to a dictionary.

get_hash
handle_unexpected_job_state
job_to_json
register_job_id
submit_job

classmethod deserialize(job_dict, machine=None)

convert the job_dict to a Submission class object

Parameters

submission_dictdict

path-like, the base directory of the local tasks

Returns

submissionJob

the Job class instance converted from the job_dict

get_hash()

get_job_state()

get the jobs. Usually, this method will query the database of slurm or pbs job scheduler system and get the results.

Notes

this method will not submit or resubmit the jobs if the job is unsubmitted.

handle_unexpected_job_state()

job_to_json()

register_job_id(job_id)

serialize(if_static=False)

convert the Task class instance to a dictionary.

Parameters

if_staticbool

whether dump the job runtime infomation (job_id, job_state, fail_count, job_uuid etc.) to the dictionary.

Returns

task_dictdict

the dictionary converted from the Task class instance

submit_job()

class dpdispatcher.submission.Resources(number_node, cpu_per_node, gpu_per_node, queue_name, group_size, *, custom_flags=[], strategy={'if_cuda_multi_devices': False, 'ratio_unfinished': 0.0}, para_deg=1, module_unload_list=[], module_purge=False, module_list=[], source_list=[], envs={}, wait_time=0, **kwargs)

Bases: object

Resources is used to describe the machine resources we need to do calculations.

Parameters

number_nodeint

The number of node need for each job.

cpu_per_nodeint

cpu numbers of each node.

gpu_per_nodeint

gpu numbers of each node.

queue_namestr

The queue name of batch job scheduler system.

group_sizeint

The number of tasks in a job.

custom_flagslist of Str

The extra lines pass to job submitting script header

strategydict

strategies we use to generation job submitting scripts. if_cuda_multi_devices : bool

If there are multiple nvidia GPUS on the node, and we want to assign the tasks to different GPUS. If true, dpdispatcher will manually export environment variable CUDA_VISIBLE_DEVICES to different task. Usually, this option will be used with Task.task_need_resources variable simultaneously.

ratio_unfinishedfloat

The ratio of jobs that can be unfinished.

para_degint

Decide how many tasks will be run in parallel. Usually run with strategy[‘if_cuda_multi_devices’]

source_listlist of Path

The env file to be sourced before the command execution.

wait_timeint

The waitting time in second after a single task submitted. Default: 0.

Methods

arginfo
deserialize
load_from_dict
load_from_json
serialize

static arginfo(detail_kwargs=True)

classmethod deserialize(resources_dict)

classmethod load_from_dict(resources_dict)

classmethod load_from_json(json_file)

serialize()

class dpdispatcher.submission.Submission(work_base, machine=None, resources=None, forward_common_files=[], backward_common_files=[], *, task_list=[])

Bases: object

A submission represents a collection of tasks. These tasks usually locate at a common directory. And these Tasks may share common files to be uploaded and downloaded.

Parameters

work_basePath

the base directory of the local tasks. It is usually the dir name of project .

machineMachine

machine class object (for example, PBS, Slurm, Shell) to execute the jobs. The machine can still be bound after the instantiation with the bind_submission method.

resourcesResources

the machine resources (cpu or gpu) used to generate the slurm/pbs script

forward_common_fileslist

the common files to be uploaded to other computers before the jobs begin

backward_common_fileslist

the common files to be downloaded from other computers after the jobs finish

task_listlist of Task

a list of tasks to be run.

Methods

bind_machine(machine)	bind this submission to a machine.
check_all_finished()	check whether all the jobs in the submission.
deserialize(submission_dict[, machine])	convert the submission_dict to a Submission class object
generate_jobs()	After tasks register to the self.belonging_tasks, This method generate the jobs and add these jobs to self.belonging_jobs.
handle_unexpected_submission_state()	handle unexpected job state of the submission.
run_submission(*[, exit_on_submit, clean])	main method to execute the submission.
serialize([if_static])	convert the Submission class instance to a dictionary.
update_submission_state()	check whether all the jobs in the submission.

check_ratio_unfinished
clean_jobs
download_jobs
get_hash
register_task
register_task_list
remove_unfinished_jobs
submission_from_json
submission_to_json
try_recover_from_json
upload_jobs

bind_machine(machine)

bind this submission to a machine. update the machine’s context remote_root and local_root.

Parameters

machineMachine

the machine to bind with

check_all_finished()

check whether all the jobs in the submission.

Notes

This method will not handle unexpected job state in the submission.

check_ratio_unfinished(ratio_unfinished)

clean_jobs()

classmethod deserialize(submission_dict, machine=None)

convert the submission_dict to a Submission class object

Parameters

submission_dictdict

path-like, the base directory of the local tasks

Returns

submissionSubmission

the Submission class instance converted from the submission_dict

download_jobs()

generate_jobs()

After tasks register to the self.belonging_tasks, This method generate the jobs and add these jobs to self.belonging_jobs. The jobs are generated by the tasks randomly, and there are self.resources.group_size tasks in a task. Why we randomly shuffle the tasks is under the consideration of load balance. The random seed is a constant (to be concrete, 42). And this insures that the jobs are equal when we re-run the program.

get_hash()

handle_unexpected_submission_state()

handle unexpected job state of the submission. If the job state is unsubmitted, submit the job. If the job state is terminated (killed unexpectly), resubmit the job. If the job state is unknown, raise an error.

register_task(task)

register_task_list(task_list)

remove_unfinished_jobs()

run_submission(*, exit_on_submit=False, clean=True)

main method to execute the submission. First, check whether old Submission exists on the remote machine, and try to recover from it. Second, upload the local files to the remote machine where the tasks to be executed. Third, run the submission defined previously. Forth, wait until the tasks in the submission finished and download the result file to local directory. if exit_on_submit is True, submission will exit.

serialize(if_static=False)

convert the Submission class instance to a dictionary.

Parameters

if_staticbool

whether dump the job runtime infomation (like job_id, job_state, fail_count) to the dictionary.

Returns

submission_dictdict

the dictionary converted from the Submission class instance

classmethod submission_from_json(json_file_name='submission.json')

submission_to_json()

try_recover_from_json()

update_submission_state()

check whether all the jobs in the submission.

Notes

this method will not handle unexpected (like resubmit terminated) job state in the submission.

upload_jobs()

class dpdispatcher.submission.Task(command, task_work_path, forward_files=[], backward_files=[], outlog='log', errlog='err')

Bases: object

A task is a sequential command to be executed, as well as the files it depends on to transmit forward and backward.

Parameters

commandStr

the command to be executed.

task_work_pathPath

the directory of each file where the files are dependent on.

forward_fileslist of Path

the files to be transmitted to remote machine before the command execute.

backward_fileslist of Path

the files to be transmitted from remote machine after the comand finished.

outlogStr

the filename to which command redirect stdout

errlogStr

the filename to which command redirect stderr

Methods

deserialize(task_dict)

convert the task_dict to a Task class object

arginfo
get_hash
load_from_dict
load_from_json
serialize

static arginfo()

classmethod deserialize(task_dict)

convert the task_dict to a Task class object

Parameters

task_dictdict

the dictionary which contains the task information

Returns

taskTask

the Task class instance converted from the task_dict

get_hash()

classmethod load_from_dict(task_dict: dict) → Task

classmethod load_from_json(json_file)

serialize()

dpdispatcher.utils module

dpdispatcher.utils.generate_totp(secret: str, period: int = 30, token_length: int = 6) → int

Generate time-based one time password (TOTP) from the secret.

Some HPCs use TOTP for two-factor authentication for safety.

Parameters

secret: str

The encoded secret provided by the HPC. It’s usually extracted from a 2D code and base32 encoded.

period: int, default=30

Time period where the code is valid in seconds.

token_length: int, default=6

The token length.

Returns

token: int

The generated token.

References

https://github.com/lepture/otpauth/blob/49914d83d36dbcd33c9e26f65002b21ce09a6303/otpauth.py#L143-L160

dpdispatcher.utils.get_sha256(filename)

Get sha256 of a file.

Parameters

filename: str

The filename.

Returns

sha256: str

The sha256.

dpdispatcher.utils.rsync(from_file: str, to_file: str)

Call rsync to transfer files.

Parameters

from_file: str

SRC

to_file: str

DEST

Raises

RuntimeError

when return code is not 0

dpdispatcher.utils.run_cmd_with_all_output(cmd, shell=True)

Running the DeePMD-kit on the Expanse cluster

Expanse is a cluster operated by the San Diego Supercomputer Center. Here we provide an example to run jobs on the expanse.

The machine parameters are provided below. Expanse uses the SLURM workload manager for job scheduling. remote_root has been created in advance. It’s worth metioned that we do not recommend to use the password, so SSH keys are used instead to improve security.

{
  "batch_type": "Slurm",
  "local_root": "./",
  "remote_root": "/expanse/lustre/scratch/njzjz/temp_project/dpgen_workdir",
  "clean_asynchronously": true,
  "context_type": "SSHContext",
  "remote_profile": {
    "hostname": "login.expanse.sdsc.edu",
    "username": "njzjz",
    "port": 22
  }
}

Expanse’s standard compute nodes are each powered by two 64-core AMD EPYC 7742 processors and contain 256 GB of DDR4 memory. Here, we request one node with 32 cores and 16 GB memory from the shared partition. Expanse does not support --gres=gpu:0 command, so we use custom_gpu_line to customize the statement.

{
  "number_node": 1,
  "cpu_per_node": 1,
  "gpu_per_node": 0,
  "queue_name": "shared",
  "group_size": 1,
  "custom_flags": [
    "#SBATCH -c 32",
    "#SBATCH --mem=16G",
    "#SBATCH --time=48:00:00",
    "#SBATCH --account=rut149",
    "#SBATCH --requeue"
  ],
  "source_list": [
    "activate /home/njzjz/deepmd-kit"
  ],
  "envs": {
    "OMP_NUM_THREADS": 4,
    "TF_INTRA_OP_PARALLELISM_THREADS": 4,
    "TF_INTER_OP_PARALLELISM_THREADS": 8,
    "DP_AUTO_PARALLELIZATION": 1
  },
  "batch_type": "Slurm",
  "kwargs": {
    "custom_gpu_line": "#SBATCH --gpus=0"
  }
}

The following task parameter runs a DeePMD-kit task, forwarding an input file and backwarding graph files. Here, the data set will be used among all the tasks, so it is not included in the forward_files. Instead, it should be included in the submission’s forward_common_files.

{
    "command": "dp train input.json && dp freeze && dp compress",
    "task_work_path": "model1/",
    "forward_files": [
      "input.json"
    ],
    "backward_files": [
      "frozen_model.pb",
      "frozen_model_compressed.pb"
    ],
    "outlog": "log",
    "errlog": "err"
}

Running Gaussian 16 with failure allowed

Typically, a task will retry three times if the exit code is not zero. Sometimes, one may allow non-zero code. For example, when running large amounts of Gaussian 16 single-point calculation tasks, some of the Gaussian 16 tasks may throw SCF errors and return a non-zero code. One can append ||: to the command:

{
    "command": "g16 < input > output ||:",
    "task_work_path": "p1/",
    "forward_files": [
      "input"
    ],
    "backward_files": [
      "output"
    ]
}

This command ensures the task will always provide zero code.

Running multiple MD tasks on a GPU workstation

In this example, we are going to show how to run multiple MD tasks on a GPU workstation. This workstation does not install any job scheduling packages installed, so we will use Shell as batch_type.

{
  "batch_type": "Shell",
  "local_root": "./",
  "remote_root": "/data2/jinzhe/dpgen_workdir",
  "clean_asynchronously": true,
  "context_type": "SSHContext",
  "remote_profile": {
    "hostname": "mandu.iqb.rutgers.edu",
    "username": "jz748",
    "port": 22
  }
}

The workstation has 48 cores of CPUs and 8 RTX3090 cards. Here we hope each card runs 6 tasks at the same time, as each task does not consume too many GPU resources. Thus, strategy/if_cuda_multi_devices is set to true and para_deg is set to 6.

{
  "number_node": 1,
  "cpu_per_node": 48,
  "gpu_per_node": 8,
  "queue_name": "shell",
  "group_size": 0,
  "strategy": {
    "if_cuda_multi_devices": true
  },
  "source_list": [
    "activate /home/jz748/deepmd-kit"
  ],
  "envs": {
    "OMP_NUM_THREADS": 1,
    "TF_INTRA_OP_PARALLELISM_THREADS": 1,
    "TF_INTER_OP_PARALLELISM_THREADS": 1
  },
  "para_deg": 6
}

Note that group_size should be set to 0 (means infinity) to ensure there is only one job and avoid running multiple jobs at the same time.

Indices and tables

• Index

• Module Index

• Search Page