Skip to content

Latest commit

 

History

History
177 lines (158 loc) · 10.4 KB

README.md

File metadata and controls

177 lines (158 loc) · 10.4 KB
Raw

Deployment and service details documentation: DEPLOYMENT.md

ElasticSearch upload for CMS and HTCondor data

This package contains a set of scripts that assist in uploading data from a HTCondor pool to ElasticSearch. It queries for both historical and current job ClassAds, converts them to a JSON document, and uploads them to the local ElasticSearch instance.

It will ignore the jobs with the value DONOTMONIT in the classsad CMS_Type.

The majority of the logic is in the conversion of ClassAds to values that are immediately useful in Kibana / ElasticSearch. A significant portion of the logic is very specific to the CMS global pool - but ultimately could be reused for other HTCondor pools.

The data can currently be accessed at es-cms.cern.ch and via CERN Monit at monit-kibana.cern.ch (with index pattern monit_prod_condor_raw_metric_v002*). To be able to save visualizations on es-cms you need to access the rw version at es-cms.cern.ch/kibana_rw and be a member of the 'cms-comp-ops' CERN e-group.

Aggregated data is managed by MONIT team and can be accessed at es-monit.cern.ch and via monit-kibana-acc.cern.ch (with index pattern monit_prod_condor_agg_metric-*) . Code base of aggregated data can be found here : MONIT Gitlab Repository

Important Attributes

This service converts the raw ClassAds into JSON documents; one JSON document per job in the system. While most attributes are copied over verbatim, a few new ones are computed on insertion; this allows easier construction of ElasticSearch queries. This section documents the most commonly used and most useful attributes, along with their meaning.

Generic attributes:

  • RecordTime: When the job exited the queue or when the JSON document was last updated, whichever came first. Use this field for time-based queries.
  • GlobalJobId: Use to uniquely identify individual jobs.
  • BenchmarkJobDB12: An estimate of the per-core performance of the machine the job last ran on, based on the DB12 benchmark. Higher is better.
  • CoreHr: The number of core-hours utilized by the job. If the job lasted for 24 hours and utilized 4 cores, the value of CoreHr will be 96. This includes all runs of the job, including any time spent on preempted instance.
  • CpuTimeHr: The amount of CPU time (sum of user and system) attributed to the job, in hours.
  • CommittedCoreHr: The core-hours only for the last run of the job (excluding preempted attempts).
  • QueueHrs: Number of hours the job spent in queue before last run.
  • WallClockHr: Number of hours the job spent running. This is invariant of the number of cores; most users will prefer CoreHr instead.
  • CpuEff: The total scheduled CPU time (user and system CPU) divided by core hours, in percentage. If the job lasted for 24 hours, utilized 4 cores, and used 72 hours of CPU time, then CpuEff would be 75.
  • CpuBadput: The badput associated with a job, in hours. This is the sum of all unsuccessful job attempts. If a job runs for 2 hours, is preempted, restarts, then completes successfully after 3 hours, the CpuBadput is 2.
  • MemoryMB: The amount of RAM used by the job.
  • Processor: The processor model (from /proc/cpuinfo) the job last ran on.
  • RequestCpus: Number of cores utilized by the job.
  • RequestMemory: Amount of memory requested by the job, in MB.
  • ScheddName: Name of HTCondor schedd where the job ran.
  • Status: State of the job; valid values include Completed, Running, Idle, or Held. Note: due to some validation issues, non-experts should only look at completed jobs.
  • x509userproxysubject: The DN of the grid certificate associated with the job; for CMS jobs, this is not the best attribute to use to identify a user (prefer CRAB_UserHN).
  • DB12CommittedCoreHr, DB12CoreHr, DB12CpuTimeHr: The job's CommittedCoreHr, CoreHr, and CpuTimeHr values multiplied by the BenchmarkJobDB12 score.

CMS-specific attributes:

  • Campaign: The campaign this job belongs to; derived from the WMAgent workflow name (so, this is an approximation; may have nonsense values for strangely-named workflows). Only present for production jobs.
  • Workflow: Human-readable workflow name. Example: HIG-RunIISpring16DR80-01026_0
  • WMAgent_RequestName: The WMAgent request name (for production jobs only); example: pdmvserv_task_HIG-RunIISpring16DR80-01026__v1_T_160530_083522_4482.
  • WMAgent_SubTaskName: The WMAgent subtask name (for production jobs only); example: /pdmvserv_task_HIG-RunIISpring16DR80-01026__v1_T_160530_083522_4482/HIG-RunIISpring16DR80-01026_0
  • CMSGroups: Name of the CMS group associated with this request (production only). Example: HIG.
  • Type: the kind of job run; analysis or production.
  • TaskType: A more detailed task type classification, based on the CMSSW config. Typically, analysis, DIGI, RECO, DIGI-RECO, GEN-SIM, or Cleanup. Is set to the CMS_TaskType classad if exists, otherwise it is infered for production jobs.
  • CMS_Pool: Condor pool to which the scheduler belongs. e.g. Global, Volunteer, Tier0,
  • MegaEvents: The number of events processed by the job, in millions.
  • KEvents: The number of events processed by the job, in thousands.
  • CMSSWKLumis: The number of lumi sections processed by the job, in thousands.
  • ChirpCMSSWMaxEvents, ChirpCMSSWMaxFiles, ChirpCMSSWMaxLumis: the maximum number of events, files, and lumis (respectively) a job will process before exiting. This may not be known for all jobs; in that case, -1 is reported. These can be used to estimate the percent completion of jobs; the veracity of these attributes are not known.
  • ExitCode: Exit code of the job. If available, this is actually the exit code of the last CMSSW step.
  • CRAB_AsyncDest: The output destination CMS site name for a CRAB3 job.
  • DESIRED_CMSDataset: The primary input dataset name.
  • CRAB_DataBlock: The primary input block name. For CRAB3 jobs only.
  • DESIRED_Sites: The list of sites the job potentially could run on.
  • DataLocations: The list of known primary input dataset locations.
  • CMSSWWallHrs: The number of wall hours reported by cmsRun.
  • StageOutHrs: An estimate of the stageout time, in hours. Calculated from WallClockHr-CMSSWWallHrs. The veracity of this attribute is not known.
  • InputData: Onsite if Site is in the DataLocations list; Offsite otherwise.
  • OutputGB: Amount of output written by the CMSSW process, in gigabytes.
  • InputGB: Amount of data read by the CMSSW process; in gigabytes.
  • ReadTimeHrs and ReadTimeMins: Amount of time CMSSW spent in its IO subsystem for reads (in hours and minutes, respectively).
  • ReadOpsPercent: percentage of reads done via a single read operation (as opposed to a vectored readv operation). If a job issues a single read and a readv of 9 segments, this value would be 50.
  • ReadOpSegmentPercent: percentage of read segments done via a single read operation. If a job issues a single read and a readv of 9 segments, this value would be 10.
  • ChirpCMSSWReadOps: number of read operations performed (excludes readv activity).
  • Site: CMS site where the job ran (for example, T2_CH_CSCS).
  • Country: Country where the job ran (for example, CH).
  • Tier: Tier where the job ran (for example, T2).
  • CRAB_UserHN: for analysis jobs, CMS username of user that submitted the task.
  • CRAB_Workflow: for analysis jobs, the CRAB task name. It is of the form 170406_201711:ztu_crab_multicrab_CMEandMixedHarmonics_Pbp_HM185_250_q3_v3_185_250_1_3; for the Workflow attribute, this is shortened to ztu_crab_multicrab_CMEandMixedHarmonics_Pbp_HM185_250_q3_v3_185_250_1_3. CRAB_Workflow is considered unique for a single CRAB task; depending on the user's naming scheme, there may multiple CRAB tasks per Workflow.
  • EventRate, CpuEventRate: The number of events per second (or per CPU second) per core.
  • TimePerEvent, CpuTimePerEvent: The inverse of EventRate and CpuEventRate, respectively.
  • HasSingularity: Set to true if the job was run inside a Singularity container.

In general, the data from CMSSW itself (number of events, data read and written) may be missing. This information depends on a sufficiently recent version of CMSSW and HTCondor.

Configuration

usage: spider_cms.py [-h] [--process_queue] [--feed_es] [--feed_es_for_queues]
                     [--feed_amq] [--schedd_filter SCHEDD_FILTER]
                     [--skip_history] [--read_only] [--dry_run]
                     [--max_documents_to_process MAX_DOCUMENTS_TO_PROCESS]
                     [--keep_full_queue_data]
                     [--amq_bunch_size AMQ_BUNCH_SIZE]
                     [--es_bunch_size ES_BUNCH_SIZE]
                     [--query_queue_batch_size QUERY_QUEUE_BATCH_SIZE]
                     [--upload_pool_size UPLOAD_POOL_SIZE]
                     [--query_pool_size QUERY_POOL_SIZE]
                     [--es_hostname ES_HOSTNAME] [--es_port ES_PORT]
                     [--es_index_template ES_INDEX_TEMPLATE]
                     [--log_dir LOG_DIR] [--log_level LOG_LEVEL]
                     [--email_alerts EMAIL_ALERTS] [--collectors COLLECTORS]
                     [--collectors_file COLLECTORS_FILE]
                     [--history_query_max_n_minutes 12*60]
                     [--mock_cern_domain]

The collectors file is a json file with the pools and a list of the collectors:

{
    "Global":[ "thefirstcollector.example.com:8080","thesecondcollector.other.com"],
    "Volunteer":["next.example.com"],
    "ITB":["newcollector.virtual.com"]
}

Example of the command:

# To process the condor queues without sending the documents to es-cms
python spider_cms.py --log_dir $LOGDIR --log_level WARNING --feed_amq\
                     --email_alerts '[email protected]'\
                     --skip_history --process_queue\
                     --query_queue_batch_size 100 --query_pool_size 16\
                     --upload_pool_size 8\
                     --collectors_file $WORKDIR/etc/collectors.json