Grid Computing Status

1
Grid Computing Status

• Michael P. McCumber
• Task Force Meeting
• April 3, 2006

2
Grid Operational Status
RCF
SBC
Submit from
CCJ, UNM, VAN
RCF
SBC
Process at
CCJ, UNM, VAN
dCache
SBC
Return to
RCF
(SRM dcache_ferry)
(dotted returns - globus-url-copy needed
cluster-wide or additional SRM-like
implementation on grid-host)
(green submit) working pre-upgrade
3
Current Grid Job/SRM Job Logistics
Local Grid Host
Pacman Repository
Remote Grid Host
Remote Node
scratch-dir
semaphore-dir
output-dir
To remove execution cluster dependence, these
directories have been standardized.
dCache
4
Grid Job XML Layout
ltxml tag defines XML version /gt ltjob tag sets
input file number gt ltcommand taggt
call pacman source environment
link output to HOME/dcache_ferry/transfer
set seeds/initialize code execute
code create semaphore lt/commandgt
ltinput tag sets input file /gt ltstdout tag
sets output log file/gt ltstderr tag sets
error log file /gt lt/jobgt
PJS will generate this automatically for
simulation users.
Scratch directory available in SCRATCH
5
dCache_ferry (SRM job)
As before a crontab job checks for semaphore
files, and runs an SRM transfer. The cycle is
currently throttled to 5 mins to prevent local
disk quota overruns. To the same end, Ive
designed the transfers to delete local copies
once a dcache transfer has been achieved. Also
since running on the remote cluster without user
login is required, Ive re-designed SRM to open a
grid-proxy when file transfers are needed. (Put
proxy password in /srm/bin/open_proxy.csh) Could
be run by a ferry captain with world writable
directories
Required directory structure
HOME
dcache_ferry
semaphore
todo
doing
done
transfer
6
Raising a Semaphore

• Once output has been sent to dcache_ferry/transfer
  directory, a semaphore is made in semaphore/todo
  such that
• 1) The file has a unique name (usually based
  on JobID, run number, etc)
• 2) The file contains a line specifying the
  full path of the file to be transferred and the
  destination directory in dCache, seperated by a
  space.
• Ex
• /phenix/u/mccumber/dcache_ferry/transfer/exampleOu
  tput.root /pnfs/rcf.bnl.gov/user/mccumber/someDire
  ctory

7
Current Available Policies

• Usage phenix-grid-submit p ltPolicy Namegt
  ltjobXMLgt
• Policies
• Grid_Policy (contains all queues)
• Sim_Policy (general queues simulation queues)
• BNL_Only_Policy
• SBC_Only_Policy
• RCAS_Only_Policy
• RCRS_Only_Policy

8
Current Development Issue
Old design makes pointless and cumbersome grid
transfers
Old Structure
Exodus
PISA
Reco
RCF
RCF
dCache
New Structure
Exodus
PISA
Reco
dCache
Still solving internal conflicts
9
Other Current/Future Spearheads

• Diagnose and solve upgrade issues
• Updating changes to PJS in CVS and Grid
  documentation
• Incorporate into grid, clusters at Vanderbilt
  (positive response), UNM CCJ (died in Limbo,
  waiting for requested access)
• Solve remote-transfer grid-proxy management
  problem (Done)
• Add option of non-dCache return of output
  (Currently not feasible)
• Need automated grid-monitoring software? Could be
  done via dynamic production of globalConfig.xml
  based on open path ways? Ferry Captain?
• Simplify, simplify, simplify