CVMFS:` Software`Access`Anywhere`
Contents
1. Accessing Multiple Repositories e Not efficient in current implementation Considered an experimental feature Disallowed by default But should be ok for occasional switching from one repository to another say lt 0 1 Hz e To enable multi repository access in a single parrot session export PARROT_ALLOVV _SWITCHING_CVMFS_REPOSITORIES 1 Accessing Other Repositories e By default Parrot knows about the CERN repositories e Can configure Parrot to access other repositories export PARROT_CVMFS_REPO cms hep wisc edu pubkey path to cms hep wisc edu pub url http cvmfsO 7 hep wisc edu cvmts cms hep wisc edu Or use equivalent parrot_run r option e See Parrot user s manual for more cvmfs options e g local cache quota Use case FUSE CVMFS at home glidein Parrot abroad dea Job can expect uniform CVMFS access wherever it lands No need to modify job code for different environments e Campus machines we administer e OSG machines we don t administer A glideinWMS Job Wrapper If job says it requires CVMFS Wraps job in parrot Uses site squid if possible Otherwise need a friendly squid at home e May limit scalability e Access control See httos github com dcbradle parrot glidein wrapper slideinWMS CVMIFS local cache e Two cases Using glexec e Each job has its own disk cache e Deleted when job exits Not using glexec e Cache is saved for li2. atlas cern ch etc e OASIS OSG project under development VOs may publish files in repository hosted by OSG Alternative to maintaining files in OSG APP at all target sites e Wisconsin OSG APP GLOW OSG site VOs write to it like any other OSG APP CVMES Server e Only needed if you wish to create your own repository e Lightweight service Kernel module to detect updates Program to prepare published files httpd to serve files Most I O done by proxies e May also want a mirror server httpd periodic sync of repository files Managing the Repository Simple case one software maintainer cvmfs user Updates software tree Triggers cvmfs publication step New files show up on clients an hour later or less Managing the Repository e More complicated scenario implementing OSG APP with CVMFS There are many software maintainers We don t want them to have to trigger publication e Tried periodically running publication Caused long delays and or write errors to software maintainers operating at time of publication e Instead using periodic rsync from user maintained tree into cvmfs software tree Then publish to cvmfs Software maintainers are never blocked Wisconsin OSG APP Repository cvmtsO1 hep wisc edu CE osggrid01 hep wisc edu nfs mount writeable gt cvmfs cms hep wisc edu tree cvmfs pub cms hep wisc edu worker node CVME
3. CVIVMIFS Software Access Anywhere Dan Bradley dan hep wisc edu Any data Any time Anywhere Project Outline Benefits of CVMFS to campus grid Installing FUSE client Using Parrot client non root A glideinWMSsSS plugin Existing repositories Hosting your own repository Some best practices What is CVMFS e Network filesystem e Read only POSIX interface cvmfs MyRepo FUSE mounted e Fetches files via http Verifies data integrity e Aggressive caching Local disk Web proxies Dan Bradley Benefits of CVMFS to Campus Grids e Well suited for software distribution Easily scalable e Local disk cache for repeated access e Add more web proxies as needed Highly available e Robust error handling failover offline mode e Add more repository mirrors as needed Secure access over untrusted networks e Strong security mechanisms for data integrity Works across administrative domains e Including unprivileged environments Parrot Truth in Advertising Young project Active development Small team Set expectations accordingly e g server component rarely used outside CERN SO more rough edges than client which is used by many LHC sites RP WN e Getting the FUSE Client Install rom Tell it which http proxies to use Allocate cache space Enable desired repositories Installing FUSE Client e RPMs are available from CERN and OSG e CERN htt
4. FES chadow tree gt cvmfs cms hep wisc edu cache01 OSG glidein cvmfs03 hep wisc edu cvmfs_server publish parrot cvmfs cms hep wisc edu local cache Some CVMES Best Practices e Following examples are for HTCondor Ideas are more general Integrating with HTCondor health check e Problem job runs and fails on machine with broken CVMFS e g cache is on broken full filesystem e How to avoid such black holes startd cron job tests for working cvmfs Publishes MyRepo CVMFS Exists True e Actual expression iffhenElse isUndefined LastHeardFrom CurrentTime LastHeardFrom 1352866188 lt 3600 e True until test expires in 1 hour Job requires TARGET MyRepo CVMEFS_ Exists True check cvmfs startd cron job See https github com dcbradley startd cron Basic functional test Monitor cache space e Important if cache does not have its own dedicated partition Advertise current CVMFS catalog version Integration with HTCondor Stale FS e Problem job runs and fails on machine that does not yet see latest cvmfs contents e How to avoid this race condition startd cron job publishes catalog version MyRepo CVMEFS Revision 4162 Job should require execute node revision gt submit node revision e For OSG jobs we do this in condor pm Example Job Set the following to output of command attr q g revision cvmfs myrepo SubmitNodeC
5. VMFSRevision 1234 Requirements TARGET MyRepo_CVMFS_ Exists amp amp TARGET MyRepo CVMFS_ Revision gt SubmitNodeCVMFSRevison Links CVMES website http cernvm cern ch portal filesystem Parrot website http www nd edu ccl software parrot Parrot CVMFS job wrapper for glideinWMS httos github com dcbradle parrot glidein wrapper CVMFS OSG APP implementation https github com dcbradley cvmfs osg a HTCondor cvmfs startd cron script httos github com dcbradley startd cron
6. fespan of glidein e May improve efficiency for very short jobs e Do we need glexec Wrapper uses Parrot s identity boxing feature e Provides privilege separation between job and glidein e But cannot be 100 trusted yet due to wrapper running in user controlled environment work in progress slideinWMS parrot cfg Configure parrot cvmfs options Here we just set the local cache quota Only default CERN repositories are enabled here PARROT _CVMFS_REPO lt default repositories gt quota_limit 4000 guota_threshold 2000 Central proxies to use for CVMIFS if the local site proxy cannot be used CVMFS_PROXIES http cacheO1 example edu 3001 http cacheO02 example 3001 CVMEFS repository to use to test site web proxy CVMFS_TEST_REPO http cvmfs stratum one cern ch opt cms path to test to validate cvmfs access CVMFS_TEST_PATH cvmfs cms cern ch If true and parrot can t access CVMFS_TEST_PATH abort glidein startup GlideinRequiresParrotCVMFS false If true all jobs are wrapped with parrot regardless of job s RequireCVMFS attribute GlideinAlwaysUseParrotWrapper false Example glideinWMS job tell glidein to wrap the job In parrot only relevant if glidein config makes this feature optional RequiresCVMFS True Executable my_program Output stdout Error stderr Queue Existing Repositories e CERN repositories http cernvm cern ch portal cvmfs examples erid cern ch cms cern ch
7. o cernvm cern ch portal filesystem What if I am not root e Parrot Virtual Filesystem No root privileges required Works as job wrapper parrot run cvmfs repo MyProgram See http www nd edu ccl software parrot Example Parrot Setup wget http www nd edu ccl software files cctools 3 6 1 x86_64 redhato tar gz tar xzf cctools 3 6 1 x86_64 redhat5 tar gz export PATH pwd cctools 3 6 1 x86_64 redhat5 bin 6PATH export HTTP_PROXY frontierO1 hep wisc edu 3128 parrot_run bash bash 3 2 Is cvmfs grid cern ch 3 1 22 0 3 1 39 0 3 1 41 0 3 1 45 0 3 2 11 1 default etc glite Parrot Performance Cost 14000 12000 a FUSE CVMFS For typical CMS jobs 10000 _ Parrot CVMFS running under Parrot is not much slower 4000 _ a Your mileage may ai 2000 vary r Assume 5 sim analysis 1 analysis 2 performance hit until proven otherwise Experience in CMS 8000 6000 seconds Parrot Cache e CVMFS local cache is in parrot tmp area Default tmp parrot lt uid gt Only one instance of parrot can use it at a time Override with parrot_ run t lt path gt e e g batch job could put it in per job tmp dir e Comparison to FUSE CVMFS Local cache not shared between batch slots e So uses more bandwidth and disk space If cache deleted after job runs successive jobs in same slot must start from scratch e Could be a problem for short jobs e g O 1 minute jobs
Download Pdf Manuals
Related Search