Skip to content

Commit

Permalink
new format of schedd.submit)/spool() fix dmwm#8336 fix dmwm#8333
Browse files Browse the repository at this point in the history
  • Loading branch information
@belforte
belforte committed May 30, 2024
1 parent f499fd5 commit 9caf99e
Showing 1 changed file with 124 additions and 108 deletions.
232 changes: 124 additions & 108 deletions src/python/TaskWorker/Actions/DagmanSubmitter.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,6 @@

import os
import copy
import json
import time
import sys

Expand Down Expand Up @@ -34,56 +33,56 @@
## These are the CRAB attributes that we want to add to the job class ad when
## using the submitDirect() method.
SUBMIT_INFO = [ \
('CRAB_ReqName', 'requestname'),
('CRAB_Workflow', 'workflow'),
('CMS_JobType', 'jobtype'),
('CRAB_JobSW', 'jobsw'),
('CRAB_JobArch', 'jobarch'),
('DESIRED_CMSDataset', 'inputdata'),
('CRAB_DBSURL', 'dbsurl'),
('CRAB_PublishName', 'publishname'),
('CRAB_Publish', 'publication'),
('CRAB_PublishDBSURL', 'publishdbsurl'),
('CRAB_PrimaryDataset', 'primarydataset'),
('CRAB_ISB', 'cacheurl'),
('CRAB_AdditionalOutputFiles', 'addoutputfiles'),
('CRAB_EDMOutputFiles', 'edmoutfiles'),
('CRAB_TFileOutputFiles', 'tfileoutfiles'),
('CRAB_TransferOutputs', 'saveoutput'),
('CRAB_SaveLogsFlag', 'savelogsflag'),
('CRAB_UserDN', 'userdn'),
('CRAB_UserHN', 'userhn'),
('CRAB_AsyncDest', 'asyncdest'),
#('CRAB_StageoutPolicy', 'stageoutpolicy'),
('CRAB_UserRole', 'tm_user_role'),
('CRAB_UserGroup', 'tm_user_group'),
('CRAB_TaskWorker', 'worker_name'),
('CRAB_RetryOnASOFailures', 'retry_aso'),
('CRAB_ASOTimeout', 'aso_timeout'),
('CRAB_RestHost', 'resthost'),
('CRAB_DbInstance', 'dbinstance'),
('CRAB_NumAutomJobRetries', 'numautomjobretries'),
('CRAB_SplitAlgo', 'splitalgo'),
('CRAB_AlgoArgs', 'algoargs'),
('CRAB_LumiMask', 'lumimask'),
('CRAB_JobCount', 'jobcount'),
('CRAB_UserVO', 'tm_user_vo'),
('CRAB_SiteBlacklist', 'siteblacklist'),
('CRAB_SiteWhitelist', 'sitewhitelist'),
('RequestMemory', 'tm_maxmemory'),
('RequestCpus', 'tm_numcores'),
('MaxWallTimeMins', 'tm_maxjobruntime'),
('MaxWallTimeMinsRun', 'tm_maxjobruntime'),
('MaxWallTimeMinsProbe', 'maxproberuntime'),
('MaxWallTimeMinsTail', 'maxtailruntime'),
('+CRAB_ReqName', 'requestname'),
('+CRAB_Workflow', 'workflow'),
('+CMS_JobType', 'jobtype'),
('+CRAB_JobSW', 'jobsw'),
('+CRAB_JobArch', 'jobarch'),
('+DESIRED_CMSDataset', 'inputdata'),
('+CRAB_DBSURL', 'dbsurl'),
('+CRAB_PublishName', 'publishname'),
('+CRAB_Publish', 'publication'),
('+CRAB_PublishDBSURL', 'publishdbsurl'),
('+CRAB_PrimaryDataset', 'primarydataset'),
('+CRAB_ISB', 'cacheurl'),
('+CRAB_AdditionalOutputFiles', 'addoutputfiles'),
('+CRAB_EDMOutputFiles', 'edmoutfiles'),
('+CRAB_TFileOutputFiles', 'tfileoutfiles'),
('+CRAB_TransferOutputs', 'saveoutput'),
('+CRAB_SaveLogsFlag', 'savelogsflag'),
('+CRAB_UserDN', 'userdn'),
('+CRAB_UserHN', 'userhn'),
('+CRAB_AsyncDest', 'asyncdest'),
#('+CRAB_StageoutPolicy', 'stageoutpolicy'),
('+CRAB_UserRole', 'tm_user_role'),
('+CRAB_UserGroup', 'tm_user_group'),
('+CRAB_TaskWorker', 'worker_name'),
('+CRAB_RetryOnASOFailures', 'retry_aso'),
('+CRAB_ASOTimeout', 'aso_timeout'),
('+CRAB_RestHost', 'resthost'),
('+CRAB_DbInstance', 'dbinstance'),
('+CRAB_NumAutomJobRetries', 'numautomjobretries'),
('+CRAB_SplitAlgo', 'splitalgo'),
('+CRAB_AlgoArgs', 'algoargs'),
('+CRAB_LumiMask', 'lumimask'),
('+CRAB_JobCount', 'jobcount'),
('+CRAB_UserVO', 'tm_user_vo'),
('+CRAB_SiteBlacklist', 'siteblacklist'),
('+CRAB_SiteWhitelist', 'sitewhitelist'),
('+RequestMemory', 'tm_maxmemory'),
('+RequestCpus', 'tm_numcores'),
('+MaxWallTimeMins', 'tm_maxjobruntime'),
('+MaxWallTimeMinsRun', 'tm_maxjobruntime'),
('+MaxWallTimeMinsProbe', 'maxproberuntime'),
('+MaxWallTimeMinsTail', 'maxtailruntime'),
('JobPrio', 'tm_priority'),
('CRAB_FailedNodeLimit', 'faillimit'),
('CRAB_DashboardTaskType', 'taskType'),
('CRAB_MaxIdle', 'maxidle'),
('CRAB_MaxPost', 'maxpost'),
('CMS_Type', 'cms_type'),
('CMS_WMTool', 'cms_wmtool'),
('CMS_TaskType', 'cms_tasktype'),
('+CRAB_FailedNodeLimit', 'faillimit'),
('+CRAB_DashboardTaskType', 'taskType'),
('+CRAB_MaxIdle', 'maxidle'),
('+CRAB_MaxPost', 'maxpost'),
('+CMS_Type', 'cms_type'),
('+CMS_WMTool', 'cms_wmtool'),
('+CMS_TaskType', 'cms_tasktype'),
]

def addCRABInfoToClassAd(ad, info):
Expand All @@ -110,6 +109,16 @@ def addCRABInfoToClassAd(ad, info):
ad[adName] = classad.ExprTree(str(adVal))


def addCRABInfoToJobJDL(jdl, info):
"""
given a submit objecty, add in the appropriate CRAB_& attributes
from the info directory
"""
for adName, dictName in SUBMIT_INFO:
if dictName in info and (info[dictName] is not None):
jdl[adName] = info[dictName]


class ScheddStats(dict):
""" collect statistics of submission success/failure at various schedulers """
def __init__(self):
Expand Down Expand Up @@ -475,86 +484,93 @@ def submitDirect(self, schedd, cmd, arg, info): #pylint: disable=R0201
"""
Submit directly to the schedd using the HTCondor module
"""
dagAd = classad.ClassAd()
addCRABInfoToClassAd(dagAd, info)
#dagJDL = classad.ClassAd()
#addCRABInfoToClassAd(dagJDL, info)
dagJDL = htcondor.Submit()
addCRABInfoToJobJDL(dagJDL, info)

if info["CMSGroups"]:
dagAd["CMSGroups"] = ','.join(info["CMSGroups"])
dagJDL["+CMSGroups"] = classad.quote(','.join(info["CMSGroups"]))
else:
dagAd["CMSGroups"] = classad.Value.Undefined
dagJDL["+CMSGroups"] = classad.Value.Undefined

# NOTE: Changes here must be synchronized with the job_submit in DagmanCreator.py in CAFTaskWorker
dagAd["CRAB_Attempt"] = 0
dagAd["CMS_SubmissionTool"] = "CRAB"
dagJDL["+CRAB_Attempt"] = 0
dagJDL["+CMS_SubmissionTool"] = "CRAB"
# We switched from local to scheduler universe. Why? It seems there's no way in the
# local universe to change the hold signal at runtime. That's fairly important for our
# resubmit implementation.
#dagAd["JobUniverse"] = 12
dagAd["JobUniverse"] = 7
dagAd["HoldKillSig"] = "SIGUSR1"
dagAd["X509UserProxy"] = info['user_proxy']
dagAd["Requirements"] = classad.ExprTree('true || false')
# SB April 2024 the following horrible linw could be replace in current HTConcdor by
# dagAd["Environmnet"] = "PATH=...; CRAB3_VERSION=...; ...." see
# https://htcondor.readthedocs.io/en/latest/man-pages/condor_submit.html#environment
dagAd["Environment"] = classad.ExprTree('strcat("PATH=/usr/bin:/bin CRAB3_VERSION=3.3.0-pre1 CONDOR_ID=", ClusterId, ".", ProcId," %s")' % " ".join(info['additional_environment_options'].split(";")))
dagAd["RemoteCondorSetup"] = info['remote_condor_setup']

dagAd["CRAB_TaskSubmitTime"] = info['start_time']
dagAd['CRAB_TaskLifetimeDays'] = TASKLIFETIME // 24 // 60 // 60
dagAd['CRAB_TaskEndTime'] = int(info['start_time']) + TASKLIFETIME
#dagJDL["JobUniverse"] = 12
dagJDL["JobUniverse"] = 7
dagJDL["HoldKillSig"] = "SIGUSR1"
dagJDL["X509UserProxy"] = info['user_proxy']
dagJDL["Requirements"] = "True || False"
EnvironmentString = "PATH=/usr/bin:/bin CRAB3_VERSION=3.3.0-pre1"
EnvironmentString += " CONDOR_ID=$(ClusterId).$(ProcId)"
EnvironmentString += " " + " ".join(info['additional_environment_options'].split(';'))
# Environment command in JDL requires proper quotes https://htcondor.readthedocs.io/en/latest/man-pages/condor_submit.html#environment
dagJDL["Environment"] = classad.quote(EnvironmentString)
dagJDL["+RemoteCondorSetup"] = info['remote_condor_setup']
dagJDL["+CRAB_TaskSubmitTime"] = info['start_time']
dagJDL['+CRAB_TaskLifetimeDays'] = TASKLIFETIME // 24 // 60 // 60
dagJDL['+CRAB_TaskEndTime'] = int(info['start_time']) + TASKLIFETIME
#For task management info see https://github.com/dmwm/CRABServer/issues/4681#issuecomment-302336451
dagAd["LeaveJobInQueue"] = classad.ExprTree("true")
dagAd["PeriodicHold"] = classad.ExprTree("time() > CRAB_TaskEndTime")
dagAd["TransferOutput"] = info['outputFilesString']
dagAd["OnExitHold"] = classad.ExprTree("(ExitCode =!= UNDEFINED && ExitCode != 0)")
dagAd["OnExitRemove"] = classad.ExprTree("( ExitSignal =?= 11 || (ExitCode =!= UNDEFINED && ExitCode >=0 && ExitCode <= 2))")
dagAd["OtherJobRemoveRequirements"] = classad.ExprTree("DAGManJobId =?= ClusterId")
dagAd["RemoveKillSig"] = "SIGUSR1"
dagJDL["LeaveJobInQueue"] = "True"
dagJDL["PeriodicHold"] = "time() > CRAB_TaskEndTime"
dagJDL["transfer_output_files"] = info['outputFilesString']
dagJDL["OnExitHold"] = "(ExitCode =!= UNDEFINED && ExitCode != 0)"
dagJDL["OnExitRemove"] = "( ExitSignal =?= 11 || (ExitCode =!= UNDEFINED && ExitCode >=0 && ExitCode <= 2))"
# dagJDL["OtherJobRemoveRequirements"] = "DAGManJobId =?= ClusterId" # appears unused SB
dagJDL["RemoveKillSig"] = "SIGUSR1"

# SB April 2024: the following write can likely be replaced by
# print(dagAd, file=fd) which is the same as fd.write(str(dagAd))
# print(dagJDL, file=fd) which is the same as fd.write(str(dagJDL))
# letting internal classad code do the formatting (in new format, which is fine
# since now we read with classad.parseOne which handles both old and new)
with open('subdag.ad', 'w', encoding='utf-8') as fd:
for k, v in dagAd.items():
if k == 'X509UserProxy':
v = os.path.basename(v)
if isinstance(v, str):
value = classad.quote(v)
elif isinstance(v, bytes):
# we only expect strings in the code above, but.. just in case,
# be prarped for bytes in case it requires a different handling at some point
value = classad.quote(v)
elif isinstance(v, classad.ExprTree):
value = repr(v)
elif isinstance(v, list):
value = "{{{0}}}".format(json.dumps(v)[1:-1])
else:
value = v
fd.write(f"+{k} = {value}\n")

dagAd["TaskType"] = "ROOT"
dagAd["Out"] = str(os.path.join(info['scratch'], "request.out"))
dagAd["Err"] = str(os.path.join(info['scratch'], "request.err"))
dagAd["Cmd"] = cmd
dagAd['Args'] = arg
dagAd["TransferInput"] = str(info['inputFilesString'])
print(dagJDL, file=fd)
# for k, v in dagJDL.items():
# if k == 'X509UserProxy':
# v = os.path.basename(v)
# if isinstance(v, str):
# value = classad.quote(v)
# elif isinstance(v, bytes):
# # we only expect strings in the code above, but.. just in case,
# # be prarped for bytes in case it requires a different handling at some point
# value = classad.quote(v)
# elif isinstance(v, classad.ExprTree):
# value = repr(v)
# elif isinstance(v, list):
# value = "{{{0}}}".format(json.dumps(v)[1:-1])
# else:
# value = v
# fd.write(f"+{k} = {value}\n")

dagJDL["+TaskType"] = "ROOT"
dagJDL["output"] = os.path.join(info['scratch'], "request.out")
dagJDL["error"] = os.path.join(info['scratch'], "request.err")
dagJDL["Cmd"] = cmd
dagJDL['Args'] = arg
dagJDL["transfer_input_files"] = info['inputFilesString']

htcondor.param['DELEGATE_FULL_JOB_GSI_CREDENTIALS'] = 'true'
htcondor.param['DELEGATE_JOB_GSI_CREDENTIALS_LIFETIME'] = '0'
resultAds = []
try:
clusterId = schedd.submit(dagAd, 1, True, resultAds)
schedd.spool(resultAds)
submitResult = schedd.submit(dagJDL, 1, True, resultAds)
clusterId = submitResult.cluster()
htcId = f"{clusterId}.0" # out cluster has only 1 job
resultAds = submitResult.clusterad()
myjobs = dagJDL.jobs(count=1, clusterid=clusterId)
schedd.spool(list(myjobs))
except htcondor.HTCondorException as hte:
raise TaskWorkerException(f"Submission failed with:\n{hte}") from hte

if 'ClusterId' in resultAds[0]:
htcId = f"{resultAds[0]['ClusterId']}.{resultAds[0]['ProcId']}"
schedd.edit([htcId], "LeaveJobInQueue", classad.ExprTree("true"))
else:
raise TaskWorkerException("Submission failed: no ClusterId was returned")
# if 'ClusterId' in resultAds[0]:
# htcId = f"{resultAds[0]['ClusterId']}.{resultAds[0]['ProcId']}"
# schedd.edit([htcId], "LeaveJobInQueue", classad.ExprTree("true"))
# else:
# raise TaskWorkerException("Submission failed: no ClusterId was returned")

# notice that the clusterId might be set even if there was a failure.
# e.g. if the schedd.submit succeded, but the spool call failed
Expand Down

0 comments on commit 9caf99e

Please sign in to comment.