The QCG-Accounting Agent

Architecture

QCG-Accounting internal architecture

Installation

You can install the package using the QosCosGrid yum repository:

yum install qcg-accounting qcg-accounting-logrotate

Configuration

The whole configuration of QCG-Accounting agent is stored in single properties files (/etc/qcg/qcg-acc/config.properties). List of configuration properties:

Common

  • qcg.site.name - your GOCDB site name,
  • qcg.batch.server - hostname where the batch server is running,
  • qcg.parser.plugin - the name of the log parser plugin (e.g. pbs). Delete this property if the agent has no access to LRMS logs,
  • qcg.publishers.plugins - the coma separated list of publisher plugins (e.g. bat,apel)
  • qcg.debug - if set to true produce more verbose messages,
  • qcg.state.dir - local state directory (default: /var/run/qcg/qcg-acc/),
  • qcg.max.delay - maximum random delay before reporting starts, the random delay was introduced in order to avoid sending reports by all sites in the same time,
  • qcg.subprocess.timeout - the general timeout after which child processes started by QCG-Accounting (e.g. log parsers) will be killed; the timeout value less than 0 means that the timeout will be disabled (default -1),
  • qcg.default.vo - the default VO name sent in case no FQAN was available (default: "vo.plgrid.pl"), e.g. job was submitted using non-VOMS proxy,
  • qcg.db.pass - password of the QCG-Computing database (see <Database> section of the qcg-compd.xml file),

If your database setup is not standard you may need to configure also the following properties:

  • qcg.db.host - QCG-Computing database host,
  • qcg.db.port - QCG-Computing database port,
  • qcg.db.name - QCG-Computing database name,
  • qcg.db.user - QCG-Computing database name.
  • qcg.db.max.days - Limit processed job records to those that has started N days ago (Default is to look 90 days back)
  • qcg.db.max.records - The maximum number of jobs processed in single run (default 5000)

Also if you want to report job as different (by default the QCG-Accounting agent tries to guess local hostname automatically) submit host than you may want to set the following property:

  • qcg.submit.host=host.second.alias

Parser plugins

SLURM plugin - slurm

No configuration needed. The plugin assumes that the sacct command is usable on the qcg machine.

PBS Pro and Torque log parser - pbs

  • qcg.pbs.home - the root of the Torque/PBS Pro spool directory (e.g. /var/torque).
  • qcg.pbs.max.days - max number of days to look back into the past (default: 7 days).

Publishers plugins

BAT publisher (PL-Grid only) - bat

At first you must ask the BAT administrator to provide you all credentials (username/password and X.509 certificate) needed to connect to the BAT. Copy the received keystore into the file /etc/qcg/qcg-acc/truststore.ts (make sure that this file is only readable by root).

  • qcg.bat.user and qcg.bat.pass - put here values provided by the BAT administrator
  • qcg.bat.keystore.pass - keystore pass (provided with key by the BAT administrator)
  • qcg.bat.test - enables test mode (i.e. do not send records to BAT broker) - default: false.
  • qcg.bat.grid.only - set this to true if you do not want to report LRMS specific job information.

APEL SSM publisher - apel

  • Install APEL SSM2 from the UMD-3 repository:
      yum install apel-ssm
    
  • Make sure that /var/spool/apel/outgoing/ exists:
    mkdir /var/spool/apel/outgoing/
    chmod 0700 /var/spool/apel/outgoing/
    
  • In the GOCDB add new endpoint for QCG host of gLite-APEL service type providing its Host DN (need to be able to publish records to APEL).
  • In the GOCDB add new endpoint for QCG host of APEL service type providing its Host DN (neeed to be monitored by Nagios).
  • Update the configuration of APEL SSM2 in the /etc/apel/sender.cfg file:
    • Configure APEL SSM2 to publish CPU Accounting data  instruction,
    • Provide appropriate authentication / authorization data,
    • Adjust logging. IMPORTANT: If you want to pipe output of the ssmsend command to the QCG-Accounting log file, ensure that the console output is enabled (console: true).

Then configure the plugin itself:

  • qcg.ssm.msg.dir - directory for outgoing usage record messages (default: /var/spool/apel/outgoing/),
  • qcg.ssm.benchmark.type - benchmark name: either Si2k or HEPSPEC,
  • qcg.ssm.benchmark.value - benchmark value (if cluster is composed of machines various types provide here weighted mean),
  • qcg.ssm.site.name - site name as reported to APEL (optional). Default: qcg.site.name,
  • qcg.ssm.max.records - the maximum number of records sent int single message (default: max 1000 records per file),
  • qcg.ssm.safe.mode - do not run APEL publisher if there are some old unsent records: (default: true),
  • qcg.ssm.send.timeout - the timeout after which the ssmsend program will be killed; if the value is less than 0, the timeout will be disabled; otherwise the value will overwrite the value specified in qcg.subprocess.timeout (default -1).

Known Issues - The QCG-Accounting must be installed on different machine than the glite-APEL and UNICORE SSM publishers otherwise reports may get overridden.

Grid-SAFE publisher - gridsafe

The Gird-SAFE publisher plugin was developed within the  MAPPER project to simplify gathering accounting data from many infrastructures (EGI, PRACE and campus resources). Steps needed to configure the GRID-SAFE plugin:

  • you can use the host cert-key pair to authenticate in the Grid-SAFE RUPI service, but first you need to convert it into the PKCS12 format. You must report your host DN to the Grid-SAFE administrator
    openssl pkcs12 -export -descert -inkey /etc/grid-security/hostkey.pem -in /etc/grid-security/hostcert.pem -out /etc/qcg/qcg-acc/hostcred.p12 -name "HOST Certificate"
    chmod 0400 /etc/qcg/qcg-acc/hostcred.p12
    
  • you can use the example configuration:
    qcg.gridsafe.url=https://gridsafe-mapper.drg.lrz.de:8443/axis2/services/RUPIService
    qcg.gridsafe.keystore=/etc/qcg/qcg-acc/hostcred.p12
    qcg.gridsafe.keystore.pass=gridsafepass
    qcg.gridsafe.truststore=/etc/qcg/qcg-acc/gridsafe-truststore.jks
    qcg.gridsafe.truststore.pass=storepass
    qcg.gridsafe.truststore.type=jks
    #send usage report only about the following users
    qcg.gridsafe.filter.userdn.file=http://gridsafe-mapper.drg.lrz.de/gridsafe/mapper.users
    
  • or configure it manually:
    • qcg.gridsafe.url - URL of the Grid-SAFE RUPI WebService (e.g.  https://gridsafe-mapper.drg.lrz.de:8443/axis2/services/RUPIService),
    • qcg.gridsafe.keystore - path to the keystore file for the RUPI plugin,
    • qcg.gridsafe.keystore.pass - password to access the keystore,
    • qcg.gridsafe.keystore.type - type of the keystore: pkcs12 or jks (default is pkcs12),
    • qcg.gridsafe.truststore - path to the truststore file for the RUPI plugin,
    • qcg.gridsafe.truststore.pass - password to access the truststore,
    • qcg.gridsafe.truststore.type - type of the truststore: pkcs12 or jks (default is pkcs12).

Filters

  • qcg.PUBLISHER.filter.userdn - send usage records only for jobs with the given X.509 DN
  • qcg.PUBLISHER.filter.userdn.file - send usage record only for jobs with the X.509 DN's listed in the given file (the file location can be an URL stream, e.g.  http://gridsafe-mapper.drg.lrz.de/gridsafe/mapper.users)
  • qcg.PUBLISHER.filter.project - send usage record only for jobs with the given project Id (grant)
  • qcg.PUBLISHER.filter.project.file - send usage records only for jobs with project id (grant) listed in the given file

Troubleshooting

The QCG-Accounting Agent stores all diagnostic information in the following log file: /var/log/qcg/qcg-acc/qcg-accounting.log. You may also try to set the qcg.debug configuration property to true in order to get more verbosity of log messages.

Migration from version 2.X to 3.0

  • stop cron temporary and make sure that no QCG-Accounting process is running
    /etc/init.d/crond stop
    ps -AF | grep QCGAcc
    
  • backup /opt/plgrid/var/run/qcg-acc/:
    cp -r /opt/plgrid/var/run/qcg-acc/ /opt/plgrid/var/run/qcg-acc.copy
    
  • update qcg-accounting:
    yum update qcg-accounting qcg-accounting-logrotate
    
  • copy configuration files and keystores to the new conf dir:
    cp /opt/plgrid/qcg/etc/qcg-acc/config.properties.rpmsave /etc/qcg/qcg-acc/config.properties
    cp /opt/plgrid/qcg/etc/qcg-acc/keystore.ks.rpmsave /etc/qcg/qcg-acc/keystore.ks
    
  • update any paths in config.properties (i.e. /opt/plgrid/qcg/etc/qcg-acc/ to /etc/qcg/qcg-acc/)
  • IMPORTANT copy last reported job ids:
    cp /opt/plgrid/var/run/qcg-acc.copy/* /var/run/qcg/qcg-acc/
    
  • try to run once and check for any errors (you may want to set temporary qcg.max.delay to 0):
    /usr/bin/qcg-accounting
    ...
    [    INFO] - Tue May 21 00:37:29 CEST 2013: new lastReportedID: 26957. Processing took: 0 seconds.
    
  • start again cron
    /etc/init.d/crond start
    

License

QCG-Accounting is released under the GPL license. For QosCosGrid licensing details see:  QosCosGrid license

FAQ

  • Q: I want to republish records for all jobs that started N days ago. What should I do?
  • A: You can do this by deleting /var/run/qcg/qcg-acc/PUBLISHER-NAME.last.id and setting qcg.db.max.days to the number of days back that you want to publish records. Also remember to adjust the qcg.pbs.max.days so it is not smaller than qcg.db.max.days.
  • Q: I am getting "Plugin apel throwed an exception: /var/spool/apel/outgoing51dd5ec6 message dir not empty. Please rerun ssmsend manually java.lang.IllegalStateException: /var/spool/apel/outgoing51dd5ec6 message dir not empty. Please rerun ssmsend manually" but I had run ssmsend already. What is wrong?
  • A: Some messages may be locked. Delete the lock file and run ssmsend again:
    rm /var/spool/apel/outgoing/*/*.lck
    

Release Notes

Attachments