90 likes | 204 Views
Proper use of CC at Lyon. J. Brunner. Batch farm / bqs. BQS classes. Only for test jobs !! Regular Exceptional resource needs Very long jobs Large memory/scratch Regular. Only on most powerful machines. Try to split or group your scripts so that they fit into the regular classes
E N D
Proper use of CC at Lyon J. Brunner
Batch farm / bqs • BQS classes Only for test jobs !! Regular Exceptional resource needs Very long jobs Large memory/scratch Regular Only on most powerful machines Try to split or group your scripts so that they fit into the regular classes If special queues needed : contact me
Batch farm / qsub options • Specify precisely which resources are needed: • qsub –l resource-list • t=1000000 : normalized CPU time • platform=LINUX32 : SL4 32 bit • M=512Mb : virtual memory • scratch=10000Mb : scratch size • hpss : access to HPSS • srb : use of SRB • oracle : acces of Oracle server at Lyon • u_srb_antares : Access to Antares SRB pool • u_xrootd : Use of XRootd server to acces Root files • u_sps_antares: use of Antares semiper disk space
Batch farm : HPSS/SRB • It is forbidden to • Store small files on HPSS/SRB (< 100Mbyte) • Make multiple access to the same file from many jobs • Forbidden means: jobs from users who are detected to perform operations like the ones above are blocked • Maximal 50 jobs can access in parallel SRB/Antares • Small files and multi-access files (e.g. scattering tables, noise files for TE) should be stored on $GROUP_DIR or on semiper disk space
Interactive machines • ccali / ccalisl432 • It is forbidden to • launch large scale productions and/or long jobs on these machines • To be used only for interactive work (graphics, editing) or test jobs to prepare batch productions • If observed behavior does not change, demons will be installed to kill interactive jobs which exceed certain (small) CPU time limits (e.g. 1000 sec)
How to acces/store evt files • SRB : Sput/Sget (/in2p3/mc , /in2p3/data) • HPSS : rfcp ($ANTHPSS/users) • SRB has larger buffer disks than HPSS, if you have a choice use SRB • Intermediate storage of output files • $TMPBATCH (up to 16 GBytes) • Antares semiper space • Same concept for output Root files
How to acces (read) Root files • XRootd • Acces via XRootd server ccxroot • Large buffer disk • Faster than copy from HPSS/SRB to scratch • Replace file name of input root file by • root://ccxroot:1999//hpss/in2p3.fr/group/antares/mc/neutrino/mu/l05_c09_s02/TE_3N_10pe/25920/TE.MCEW.gea.anumu_lowe_109.root
Problems with XRootd • When using with Root version earlier than 5.20 there are problems • Memory leak virtual memory limit • Corrupted files • Standard Root version at CC-Lyon 5.14 • How to use 5.20 ? • Add following lines in your .cshrc before • source $THRONG_DIR/group_cshrc • setenv ROOTSYS /usr/local/root/root_v5.20.00/root • source /usr/local/shared/bin/root_env.csh • Use antares-daq, compiled with Root 5.20 like • $ANTARES/src/DAQ_tree/Root_5.20_080902
Antares semiper disk space (new!) • So far only $GROUP_DIR (afs) available to store files which are needed by many jobs, i.e. must be visible from many machines • $GROUP_DIR total size 65 Gbytes • Now GFPS space /sps/antares 2 TBytes • Not on afs, not visible from outside • Mounted on all interactive and batch machines • Direct acces (cp) or via rfio (rfcp), e.g. from HPSS • User area /sps/antares/users • Create subdir with your login name • mkdir /sps/antares/users/brunner • Use to store small files, frequent acces files • Use for temporary analysis files • No disk quota introduced yet: Be reasonable !