SAMGrid: the merging mechanism

1
SAM-Grid the merging mechanism

• Gabriele Garzoglio for the SAM-Grid Team

2
Overview

• Context
• Montecarlo Production
• The Merging mechanism
• General comments
• Room for improvements
• Conclusion

3
Context

• JIM and SAM can be used together to produce
  montecarlo events at three execution sites
  Wisconsin, Manchester, Lyon. Other sites are also
  joining.
• For every request, montecarlo thumbnails are
  produced in parallel at each cluster.
• JIM provides a mechanism to merge the files
  produced.

4
Production mechanism

• The user submits a job to SAM-Grid to process N
  events of a certain request.
• Eventually the job is dispatched to an execution
  site.
• The grid job is split into M parallel local jobs.
• Each local job uses Runjob to produce the events
  and store the files.
• Thumbnails are stored into a SAM location (disk)
  close to the cluster.

5
Merging mechanism I

• When the production jobs are finished, a merging
  job can be submitted to SAM-Grid.
• The job needs to know a dataset definition for
  the files to be merged. Basic consistency checks
  are applied at the client site.
• Eventually the job is dispatched to an execution
  site.

6
Merging mechanism II

• A single local job is submitted to the batch
  system. The job stages the files from SAM locally
  and uses copyd0om (via Runjob) to merge them.
• merge.py manipulates the metadata and stores the
  merged files to SAM.
• When the file is handed over to FSS, the
  thumbnail locations are undeclared from the
  database.
• The files can then be physically removed

7
General Comments

• The dataset definition that identifies the files
  to merge can use dimensions like request id, grid
  job id of the production job, event size,
  dates...
• We do basic checks before submission e.g. are
  the files in the dataset unmerged thumbnails?
  AND do they have well formed parents.
• Can merge files produced in different sites the
  mechanism relies on SAM for the delivery

8
Room for improvement... I

• We dont have a way to distinguish merged
  thumbnails (no dimension, no datatier) we try
  not to depend on it and we use file names if we
  need to guess
• We should optimize the checks at the client
• We should improve the logic for storing merged
  thumbnails
• use asynchronous store instead of synchronous

9
Room for improvement... II

• Removal of used unmerged thumbnails
• using a file status vs. undeclaring the location
  from the database ?
• automatic physical file removal
• usage of sam caches (e.g. sam import file)
• Working on automation, by automatically launching
  merging jobs

10
Conclusions

• SAM-Grid uses SAM to select the files to be
  merged
• The location of production is decoupled from the
  location of merging
• We use the SAM-Grid infrastructure for job
  dispatch and monitoring
• The infrastructure is in production since March.
• The SAR institutions have developed a JIM
  interface to mc_farm and ramping up.
• We can still improve the overall robustness and
  automation