Glidein startup Internals and Glidein configuration - glideinWMS Training Jan 2012
1 like388 views
The document provides a detailed overview of the glideinWMS training at UCSD, covering the internal workings of glidein startup, configuration, and lifecycle management. It discusses various aspects such as security considerations, glidein self-termination policies, and the use of glexec for enhanced security in multi-user pilot jobs. Additionally, practical examples and configuration parameters are included for implementing glideinWMS effectively.
![Example (pseudo-)script
## check ifCMSSW installed locally
check if CMSSW installed locally
ifif [-f-f"$CMSSW" ];];then
[ "$CMSSW" then
## get the other env
get the other env
source "$CMSSW"
source "$CMSSW"
else
else
## fail validation
fail validation
echo "CMSSW not found!n" 1>&2
echo "CMSSW not found!n" 1>&2
exit 11
exit
fifi
## publish to Condor
publish to Condor
add_config_line "CMSSW_LIST" "$CMS_SW_LIST"
add_config_line "CMSSW_LIST" "$CMS_SW_LIST"
add_condor_vars_line "CMSSW_LIST" "S" "-" "+" "Y" "Y" "-"
add_condor_vars_line "CMSSW_LIST" "S" "-" "+" "Y" "Y" "-"
## if Igot here, passed validation
if I got here, passed validation
exit 00
exit
More details at http://tinyurl.com/glideinWMS/doc.prd/factory/custom_scripts.html
UCSD Jan 17th 2012 Glidein internals 14](https://image.slidesharecdn.com/gfrontend12010105glideininternals-120209211411-phpapp01/85/Glidein-startup-Internals-and-Glidein-configuration-glideinWMS-Training-Jan-2012-14-320.jpg)
Glidein startup Internals and Glidein configuration - glideinWMS Training Jan 2012
- 1. glideinWMS Training @ UCSD Glidein startup Internals and Glidein configuration by Igor Sfiligoi (UCSD) UCSD Jan 17th 2012 Glidein internals 1
- 2. Outline ● Glidein_startup internals ● Glidein lifetime ● Security considerations UCSD Jan 17th 2012 Glidein internals 2
- 3. Refresher - What is a glidein? ● A glidein is just a properly configured execution node submitted as a Grid job Central manager glidein Execution node Collector glidein Execution node Negotiator Submit node Submit node glidein Execution node Submit node Execution node glidein Schedd Startd Job UCSD Jan 17th 2012 Glidein internals 3
- 4. Refresher – Glidein startup ● glidein_startup configures and starts Condor Configure Condor G.N. Submit node Frontend node Worker node Monitor Submit node Frontend Condor glidein Central manager Startd Match CREAM Job Request glideins Factory node Condor glidein Execution node Globus Factory glidein Execution node Submit glideins UCSD Jan 17th 2012 Glidein internals 4
- 5. glidein_startup Internals UCSD Jan 17th 2012 Glidein internals 5
- 6. glidein_startup tasks ● Download scripts, parameters and Condor bins ● Validate node (environment) ● Configure Condor ● Start Condor daemon(s) ● Collect post-mortem monitoring info ● Cleanup UCSD Jan 17th 2012 Glidein internals 6
- 7. Downloading files ● Files downloaded via HTTP ● From both the factory and the frontend Web servers ● Can use local Web proxy (e.g. Squid) ● Mechanism tamper proof and cache coherent Factory node glidein_startup HTTPd ● Load files from factory Web Squid ● Load files from frontend Web Frontend node ● Run executables ● Start Condor Startd HTTPd ● Cleanup Nothing will work if HTTPd down! UCSD Jan 17th 2012 Glidein internals 7
- 8. URLs ● All URLs passed to glidein_startup as arguments ● Factory Web server ● Frontend Web server ● Squid, if any ● glidein_startup arguments defined by the factory ● Frontend Web URL passed to the Factory via request ClassAd UCSD Jan 17th 2012 Glidein internals 8
- 9. Cache coherence ● Files never change, ● There is a (set of) once uploaded to the logical to physical Web area name map files ● *_file_list.id.lst ● If the source file changes, a file with a ● The names of the list different name is files are in created on the Web ● description.id.cfg area ● The id of this list is passed as an ● Essentially argument Frontend ids fname.date passed to the to glidein_startup factory via ClassAd UCSD Jan 17th 2012 Glidein internals 9
- 10. Making the download tamper proof ● Verifying SHA1 ● The hash of the hashes signature file a Frontend hash ● Each file has a SHA1 parameter of passed to the factory associated with it glidein_startup via ClassAd ● The hashes delivered ● This is guaranteed to in a single file be secure with GRAM ● signature.id.sha1 ● List of (fname,hash) ● Description file hash pairs in the signature file ● Signature file id in the ● Still safe even if description file loaded out of order UCSD Jan 17th 2012 Glidein internals 10
- 11. Example Web area Frontend Factory Globus glidein_startup ... . .. UCSD Jan 17th 2012 Glidein internals 11
- 12. Node validation ● Run scripts / plugins provided by both factory and frontend ● The map file has a flag to distinguish scripts from “regular” files # Outfile InFile Exec/other # Outfile InFile Exec/other ######################################################### ######################################################### constants.cfg constants.b8ifnW.cfg regular constants.cfg constants.b8ifnW.cfg regular cat_consts.sh cat_consts.b8ifnW.sh exec cat_consts.sh cat_consts.b8ifnW.sh exec set_home_cms.source set_home_cms.b8ifnW.source wrapper set_home_cms.source set_home_cms.b8ifnW.source wrapper ● If a script returns with exit_code !=0, glidein_startup stops execution ● Condor never started → no user jobs ever pulled ● Will sleep for 20 mins → blackhole protection UCSD Jan 17th 2012 Glidein internals 12
- 13. Condor configuration ● Condor config values coming from files downloaded by glidein_startup ● Static from both factory and frontend config files ● More details at http://tinyurl.com/glideinWMS/doc.prd/factory/custom_vars.html ● Scripts can add, alter or delete any attribute ● Based on dynamic information ● Either discovered on the WN or by combining info from various sources ● More details at http://tinyurl.com/glideinWMS/doc.prd/factory/custom_scripts.html UCSD Jan 17th 2012 Glidein internals 13
- 14. Example (pseudo-)script ## check ifCMSSW installed locally check if CMSSW installed locally ifif [-f-f"$CMSSW" ];];then [ "$CMSSW" then ## get the other env get the other env source "$CMSSW" source "$CMSSW" else else ## fail validation fail validation echo "CMSSW not found!n" 1>&2 echo "CMSSW not found!n" 1>&2 exit 11 exit fifi ## publish to Condor publish to Condor add_config_line "CMSSW_LIST" "$CMS_SW_LIST" add_config_line "CMSSW_LIST" "$CMS_SW_LIST" add_condor_vars_line "CMSSW_LIST" "S" "-" "+" "Y" "Y" "-" add_condor_vars_line "CMSSW_LIST" "S" "-" "+" "Y" "Y" "-" ## if Igot here, passed validation if I got here, passed validation exit 00 exit More details at http://tinyurl.com/glideinWMS/doc.prd/factory/custom_scripts.html UCSD Jan 17th 2012 Glidein internals 14
- 15. Example Frontend attr ● START expression ● As GLIDECLIENT_Group_Start ● Automatically && with Factory provided one GLIDECLIENT_Group_Start = (stringListMember(GLIDEIN_SEs, DESIRED_SEs)=?=True) || (stringListMember(GLIDEIN_Site, DESIRED_Sites=?=True)) More details at http://tinyurl.com/glideinWMS/doc.prd/factory/custom_vars.html UCSD Jan 17th 2012 Glidein internals 15
- 16. Condor startup ● After validation and configuration, glidein just start condor_master ● Which in turn start condor_startd ● It is just regular Condor from here on ● Any policy the VO needs must be part of Condor configuration Central manager Submit node Execution node Collector Schedd glidein_startup Negotiator Startd UCSD Jan 17th 2012 Glidein internals 16
- 17. Post mortem monitoring ● After a glidein terminates, the logs are sent back to the Factory ● Frontend does not have access to them ● Work in progress for future glideinWMS release ● For now, ask Factory admins if debugging UCSD Jan 17th 2012 Glidein internals 17
- 18. Cleanup ● glidein_startup will remove all files before terminating ● Unless killed by the OS, of course ● Condor does something similar for the user jobs disk area after every job ● Users should not expect anything to survive their jobs UCSD Jan 17th 2012 Glidein internals 18
- 19. Glidein lifetime UCSD Jan 17th 2012 Glidein internals 19
- 20. Glidein lifetime ● Glideins are temporary resources ● Must go away after some time ● We want them to go away by their own will ● So we can monitor progress and clean up ● As opposed to being killed by Grid batch system ● Condor daemons configured to die by themselves ● Just need to tell them when UCSD Jan 17th 2012 Glidein internals 20
- 21. Limiting glidein lifetime ● Hard End-of-Life deadline ● Condor will terminate if it is ever reached ● Any jobs running at that point in time will be killed – Resulting in waste ● Deadline site specific ● Set by the Factory ● Controlled by GLIDEIN_Max_Walltime ● Condor advertises it in the glidein ClassAd ● Attribute GLIDEIN_ToDie UCSD Jan 17th 2012 Glidein internals 21
- 22. Deadline for job startup ● Starting jobs just to kill them wasteful ● Glideins set an earlier deadline for job startup ● After the deadline, Condor will terminate at job end ● Advertised as GLIDEIN_TORETIRE ● Need to know how long is the Default exists but expected job lifetime don't count on it ● Should be provided by the Frontend ● Parameter GLIDEIN_Job_Max_Time ● ToRetire=Max_Walltime-Job_Max_Time UCSD Jan 17th 2012 Glidein internals 22
- 23. Termination due to non-use ● glideins will self-terminate if not used, too ● To limit waste ● Reasons for hitting this ● No more user jobs (spikes) ● Policy problems (Frontend vs Negotiator) ● Typically uniform across the Condor pool ● Can be set by either Frontend or Factory ● Controlled by GLIDEIN_Max_Idle To give Negotiator ● Defaults to 20 mins the time to match the glidein UCSD Jan 17th 2012 Glidein internals 23
- 24. Finer grained policies ● Job_Max_Time is the typical expected job lifetime ● What if not all jobs have the same lifetime? ● Use Condor matchmaking to define finer grained policies ● To prevent long jobs from starting if they are expected to be killed before terminating ● This allows (lower priority) shorter jobs to use the CPU ● Or waste just 20 mins (instead of hours) ● Job lifetime can be difficult to know ● Even users often don't know ● Can refine after restarts (assuming they are deterministic) UCSD Jan 17th 2012 Glidein internals 24
- 25. Example START expression ● Have two different thresholds ● One for first startup, one for all following ● Useful when wide dynamic range, unpredictable GLIDECLIENT_Start = ifthenelse( LastVacateTime=?=UNDEFINED, NormMaxWallTime < (GLIDEIN_ToDie-MyCurrentTime), MaxWallTime < (GLIDEIN_ToDie-MyCurrentTime) ) UCSD Jan 17th 2012 Glidein internals 25
- 26. Other sources of waste ● Glideins must stay within the limits of the resource lease ● Typical limit is on memory usage OSG Factory defines GLIDEIN_MaxMemMBs ● If user jobs exceed that limit, the glideins may be killed by the resource provider (e.g. Grid batch system) ● Resulting in the user job being killed, too ● Thus waste UCSD Jan 17th 2012 Glidein internals 26
- 27. Example START expression ● Prevent startup of jobs that are known to use too much memory ● Unless user defines it, known only at 2nd re-run ● Also kill jobs as soon as they pass that ● So we don't have to start a new glidein for other jobs GLIDECLIENT_Start = ImageSize<=(GLIDEIN_MaxMemMBs*1024) PREEMPT = ImageSize>(GLIDEIN_MaxMemMBs*1024) More details at http://research.cs.wisc.edu/condor/manual/v7.6/10_Appendix_A.html#82447 UCSD Jan 17th 2012 Glidein internals 27
- 28. Security considerations UCSD Jan 17th 2012 Glidein internals 28
- 29. Talking to the rest of Condor ● Glideins will talk over ● Collector and glidein WAN to the rest of must whitelist each Condor other ● Strong security a must ● Schedd and glidein Central manager trusted indirectly Collector ● Security token Negotiator through Collector Submit node Execution node AN glidein_startup Schedd W Startd UCSD Jan 17th 2012 Glidein internals 29
- 30. Multi User Pilot Jobs ● A glidein typically starts with a proxy that is not representing a specific user ● Very few exceptions (e.g. CMS Organized Reco) ● Three potential security issues: ● Site does not know who the final user is (and thus cannot ban specific users, if needed) ● If 2 glideins land on the same node, 2 users may be running as the same UID (no system protection) ● User and pilot code run as the same UID (no system protection) UCSD Jan 17th 2012 Glidein internals 30
- 31. glexec ● We have one tool that can help us with all 3 ● Namely glexec ● glexec provides UID switching ● But must be installed by the Site on WNs ● Only a subset of sites have done it so far ● glexec requires a valid proxy from the final users on the submit node ● Or jobs will not match ● Not a problem for long time Grid users, like CMS, but it may be for other VOs UCSD Jan 17th 2012 Glidein internals 31
- 32. glexec in a picture Legend Central manager Pilot UID Collector User UID Negotiator Submit node Submit node Execution node Submit node glidein Schedd Startd glexec Job UCSD Jan 17th 2012 Glidein internals 32
- 33. glexec, cont ● It is in VO best interest to use glexec ● Only way to have a secure MUPJ system ● Should push sites to deploy it ● Some sites require glexec for their own interest ● To cryptographically know who the final users are ● To easily ban users May be a legal requirement ● Note on site banning ● Condor currently does not handle well glexec failures – VO admin must look for this UCSD Jan 17th 2012 Glidein internals 33
- 34. Turning on glexec ● Factory provides information about glexec installation at site ● Attribute GLEXEC_BIN = NONE | OSG | glite ● Frontend decides if it should be used ● Parameter GLIDEIN_Glexec_Use ● Possible values: – NEVER - Do not use, even if available – OPTIONAL - Use whenever available – REQUIRED - Only use sites that provide glexec UCSD Jan 17th 2012 Glidein internals 34
- 35. THE END UCSD Jan 17th 2012 Glidein internals 35
- 36. Pointers ● The official project Web page is http://tinyurl.com/glideinWMS ● glideinWMS development team is reachable at glideinwms-support@fnal.gov ● CMS AnaOps Frontend at UCSD http://glidein-collector.t2.ucsd.edu:8319/vofrontend/monitor/frontend_UCSD-v5_2/frontendStatus.html UCSD Jan 17th 2012 Glidein internals 36
- 37. Acknowledgments ● The glideinWMS is a CMS-led project developed mostly at FNAL, with contributions from UCSD and ISI ● The glideinWMS factory operations at UCSD is sponsored by OSG ● The funding comes from NSF, DOE and the UC system UCSD Jan 17th 2012 Glidein internals 37