Source code for dpdispatcher.machines.slurm

import math
import pathlib
import shlex
from typing import List

from dargs import Argument

from dpdispatcher.dlog import dlog
from dpdispatcher.machine import Machine, script_command_template
from dpdispatcher.utils.job_status import JobStatus
from dpdispatcher.utils.utils import (
    RetrySignal,
    customized_script_header_template,
    retry,
)

# from dpdispatcher.submission import Resources

slurm_script_header_template = """\
#!/bin/bash -l
#SBATCH --parsable
{slurm_nodes_line}
{slurm_ntasks_per_node_line}
{slurm_number_gpu_line}
{slurm_partition_line}"""

slurm_job_array_script_end_template = """
wait

{append_script_part}
"""


[docs] class Slurm(Machine):
[docs] def gen_script(self, job): slurm_script = super().gen_script(job) return slurm_script
[docs] def gen_script_header(self, job): resources = job.resources script_header_dict = {} script_header_dict["slurm_nodes_line"] = ( f"#SBATCH --nodes {resources.number_node}" ) script_header_dict["slurm_ntasks_per_node_line"] = ( f"#SBATCH --ntasks-per-node {resources.cpu_per_node}" ) custom_gpu_line = resources.kwargs.get("custom_gpu_line", None) if not custom_gpu_line: script_header_dict["slurm_number_gpu_line"] = ( f"#SBATCH --gres=gpu:{resources.gpu_per_node}" ) else: script_header_dict["slurm_number_gpu_line"] = custom_gpu_line if resources.queue_name != "": script_header_dict["slurm_partition_line"] = ( f"#SBATCH --partition {resources.queue_name}" ) else: script_header_dict["slurm_partition_line"] = "" if ( resources["strategy"].get("customized_script_header_template_file") is not None ): slurm_script_header = customized_script_header_template( resources["strategy"]["customized_script_header_template_file"], resources, ) else: slurm_script_header = slurm_script_header_template.format( **script_header_dict ) return slurm_script_header
@retry() def do_submit(self, job): script_file_name = job.script_file_name script_str = self.gen_script(job) job_id_name = job.job_hash + "_job_id" # script_str = self.sub_script(job_dirs, cmd, args=args, resources=resources, outlog=outlog, errlog=errlog) self.context.write_file(fname=script_file_name, write_str=script_str) script_run_str = self.gen_script_command(job) script_run_file_name = f"{job.script_file_name}.run" self.context.write_file(fname=script_run_file_name, write_str=script_run_str) # self.context.write_file(fname=os.path.join(self.context.submission.work_base, script_file_name), write_str=script_str) command = "cd {} && {} {}".format( shlex.quote(self.context.remote_root), "sbatch", shlex.quote(script_file_name), ) ret, stdin, stdout, stderr = self.context.block_call(command) if ret != 0: err_str = stderr.read().decode("utf-8") if ( "Socket timed out on send/recv operation" in err_str or "Unable to contact slurm controller" in err_str ): # server network error, retry 3 times raise RetrySignal( f"Get error code {ret} in submitting with job: {job.job_hash} . message: {err_str}" ) elif ( "Job violates accounting/QOS policy" in err_str # the number of jobs exceeds DEFAULT_MAX_JOB_COUNT (by default 10000) or "Slurm temporarily unable to accept job, sleeping and retrying" in err_str ): # job number exceeds, skip the submitting return "" raise RuntimeError( f"command {command} fails to execute\nerror message:{err_str}\nreturn code {ret}\n" ) subret = stdout.readlines() # --parsable # Outputs only the job id number and the cluster name if present. # The values are separated by a semicolon. Errors will still be displayed. job_id = subret[0].split(";")[0].strip() self.context.write_file(job_id_name, job_id) return job_id
[docs] def default_resources(self, resources): pass
@retry() def check_status(self, job): job_id = job.job_id if job_id == "": return JobStatus.unsubmitted command = 'squeue -o "%.18i %.2t" -j ' + job_id ret, stdin, stdout, stderr = self.context.block_call(command) if ret != 0: err_str = stderr.read().decode("utf-8") if "Invalid job id specified" in err_str: if self.check_finish_tag(job): dlog.info(f"job: {job.job_hash} {job.job_id} finished") return JobStatus.finished else: return JobStatus.terminated elif ( "Socket timed out on send/recv operation" in err_str or "Unable to contact slurm controller" in err_str or "Invalid user for SlurmUser" in err_str ): # retry 3 times raise RetrySignal( f"Get error code {ret} in checking status with job: {job.job_hash} . message: {err_str}" ) raise RuntimeError( f"status command {command} fails to execute." f"job_id:{job_id} \n error message:{err_str}\n return code {ret}\n" ) status_line = stdout.read().decode("utf-8").split("\n")[-2] status_word = status_line.split()[-1] if not (len(status_line.split()) == 2 and status_word.isupper()): raise RuntimeError( "Error in getting job status, " + f"status_line = {status_line}, " + f"parsed status_word = {status_word}" ) if status_word in ["PD", "CF", "S"]: return JobStatus.waiting elif status_word in ["R"]: return JobStatus.running elif status_word in ["CG"]: return JobStatus.completing elif status_word in [ "C", "E", "K", "BF", "CA", "CD", "F", "NF", "PR", "SE", "ST", "TO", ]: if self.check_finish_tag(job): dlog.info(f"job: {job.job_hash} {job.job_id} finished") return JobStatus.finished else: return JobStatus.terminated else: return JobStatus.unknown
[docs] def check_finish_tag(self, job): job_tag_finished = job.job_hash + "_job_tag_finished" return self.context.check_file_exists(job_tag_finished)
[docs] @classmethod def resources_subfields(cls) -> List[Argument]: """Generate the resources subfields. Returns ------- list[Argument] resources subfields """ doc_custom_gpu_line = "Custom GPU configuration, starting with #SBATCH" return [ Argument( "kwargs", dict, [ Argument( "custom_gpu_line", str, optional=True, default=None, doc=doc_custom_gpu_line, ) ], optional=True, doc="Extra arguments.", ) ]
[docs] def kill(self, job): """Kill the job. Parameters ---------- job : Job job """ job_id = job.job_id # -Q Do not report an error if the specified job is already completed. ret, stdin, stdout, stderr = self.context.block_call( "scancel -Q " + str(job_id) )
# we do not need to stop here if scancel failed; just continue
[docs] class SlurmJobArray(Slurm): """Slurm with job array enabled for multiple tasks in a job."""
[docs] def gen_script_header(self, job): slurm_job_size = job.resources.kwargs.get("slurm_job_size", 1) if job.fail_count > 0: # resubmit jobs, check if some of tasks have been finished job_array = set() for ii, task in enumerate(job.job_task_list): task_tag_finished = ( pathlib.PurePath(task.task_work_path) / (task.task_hash + "_task_tag_finished") ).as_posix() if not self.context.check_file_exists(task_tag_finished): job_array.add(ii // slurm_job_size) return super().gen_script_header(job) + "\n#SBATCH --array={}".format( ",".join(map(str, job_array)) ) return super().gen_script_header(job) + "\n#SBATCH --array=0-%s" % ( math.ceil(len(job.job_task_list) / slurm_job_size) - 1 )
[docs] def gen_script_command(self, job): resources = job.resources slurm_job_size = resources.kwargs.get("slurm_job_size", 1) # SLURM_ARRAY_TASK_ID: 0 ~ n_jobs-1 script_command = "case $SLURM_ARRAY_TASK_ID in\n" for ii, task in enumerate(job.job_task_list): command_env = "" command_env += self.gen_command_env_cuda_devices(resources=resources) task_tag_finished = task.task_hash + "_task_tag_finished" log_err_part = "" if task.outlog is not None: log_err_part += f"1>>{shlex.quote(task.outlog)} " if task.errlog is not None: log_err_part += f"2>>{shlex.quote(task.errlog)} " flag_if_job_task_fail = job.job_hash + "_flag_if_job_task_fail" last_err_file = job.job_hash + "_last_err_file" single_script_command = script_command_template.format( flag_if_job_task_fail=flag_if_job_task_fail, command_env=command_env, task_work_path=shlex.quote( pathlib.PurePath(task.task_work_path).as_posix() ), command=task.command, task_tag_finished=task_tag_finished, log_err_part=log_err_part, err_file=shlex.quote(task.errlog), last_err_file=shlex.quote(last_err_file), ) if ii % slurm_job_size == 0: script_command += f"{ii // slurm_job_size})\n" script_command += single_script_command script_command += self.gen_script_wait(resources=resources) script_command += "\n" if ( ii % slurm_job_size == slurm_job_size - 1 or ii == len(job.job_task_list) - 1 ): script_command += ";;\n" script_command += "*)\nexit 1\n;;\nesac\n" return script_command
[docs] def gen_script_end(self, job): # We cannot touch tag for job array # we may check task tag instead append_script = job.resources.append_script append_script_part = "\n".join(append_script) return slurm_job_array_script_end_template.format( append_script_part=append_script_part, )
@retry() def check_status(self, job): job_id = job.job_id if job_id == "": return JobStatus.unsubmitted command = 'squeue -h -o "%.18i %.2t" -j ' + job_id ret, stdin, stdout, stderr = self.context.block_call(command) if ret != 0: err_str = stderr.read().decode("utf-8") if "Invalid job id specified" in err_str: if self.check_finish_tag(job): dlog.info(f"job: {job.job_hash} {job.job_id} finished") return JobStatus.finished else: return JobStatus.terminated elif ( "Socket timed out on send/recv operation" in err_str or "Unable to contact slurm controller" in err_str ): # retry 3 times raise RetrySignal( f"Get error code {ret} in checking status with job: {job.job_hash} . message: {err_str}" ) raise RuntimeError( f"status command {command} fails to execute." f"job_id:{job_id} \n error message:{err_str}\n return code {ret}\n" ) status_lines = stdout.read().decode("utf-8").split("\n")[:-1] status = [] for status_line in status_lines: status_word = status_line.split()[-1] if not (len(status_line.split()) == 2 and status_word.isupper()): raise RuntimeError( "Error in getting job status, " + f"status_line = {status_line}, " + f"parsed status_word = {status_word}" ) if status_word in ["PD", "CF", "S"]: status.append(JobStatus.waiting) elif status_word in ["R"]: status.append(JobStatus.running) elif status_word in ["CG"]: status.append(JobStatus.completing) elif status_word in [ "C", "E", "K", "BF", "CA", "CD", "F", "NF", "PR", "SE", "ST", "TO", ]: status.append(JobStatus.finished) else: status.append(JobStatus.unknown) # running if any job is running if JobStatus.running in status: return JobStatus.running elif JobStatus.waiting in status: return JobStatus.waiting elif JobStatus.completing in status: return JobStatus.completing elif JobStatus.unknown in status: return JobStatus.unknown else: if self.check_finish_tag(job): dlog.info(f"job: {job.job_hash} {job.job_id} finished") return JobStatus.finished else: return JobStatus.terminated
[docs] def check_finish_tag(self, job): results = [] for task in job.job_task_list: task.get_task_state(self.context) results.append(task.task_state == JobStatus.finished) return all(results)
[docs] @classmethod def resources_subfields(cls) -> List[Argument]: """Generate the resources subfields. Returns ------- list[Argument] resources subfields """ doc_slurm_job_size = "Number of tasks in a Slurm job" arg = super().resources_subfields()[0] arg.extend_subfields( [ Argument( "slurm_job_size", int, optional=True, default=1, doc=doc_slurm_job_size, ), ] ) return [arg]