Review of 2012 Distributed Computing Operations
Stefan Roiser66th LHCb Week – NCB Meeting
26 November, 2012
2012 Operations 2
Content
• 2012 data processing activities• Current issues and work in progress• Processing outlook for 2013 and beyond• Conclusion
26 Nov '12
2012 Operations 3
2012 ACTIVITIES
26 Nov '12
2012 Operations 4
2012 Processing Overview(all plots of this talk since 1 Jan)
26 Nov '12
All successful Jobs of 2012
SimulationReconstruction
Reprocessing
Stripping
UserAll successful Workgrouped by Activity
CPU Efficiency ofSuccessful jobs
UserTotal
Simulation
StrippingReprocessing
Prompt Processing
LHCb was making good use of the provided resources
2012 Operations 5
Successful Work in 2012 by Site
26 Nov '12
Total 3.5 M CPU days
2012 Operations 6
MC Simulation
26 Nov '12
• Mostly running at Tier2 sites– But also at Tier0/1 when resources are available– Usually running
with lowestpriority
– Low error rate as also true for otherproduction activities
Job Characteristics• No input data• Uploading output to
“closest” storage element
2012 Operations 7
Prompt Reconstruction
26 Nov '12
• First pass reconstruction of detector data– Usually 100 % of RAW files are processed. Since
reprocessing started, only partial (~30%) reconstruction at CERN + “attached T2s”
Job Characteristics• At Tier0/1 sites• 3GB downl input from tape• ~5 GB output to tape • Job duration ~ 36 hours
2012 Operations 8
• Reprocessing of 2012 data started mid Sept.• Pushing system to its limits– Running up to 15k reconstruction jobs was a very
good stress test for post LS1 prompt processing– Data processing smooth– Hitting limits with data
movement• E.g. staging from Tape
(see Philippe’s talk)
Data Reprocessing
26 Nov '12
Job Characteristics• Same as prompt reco• + also running at Tier2 sites
2012 Operations 9
2012 Data Reprocessing
26 Nov '12
41 Tier 2 sites involved in the activity, downloading RAW file from T1 storage and providing ~ 50 % of work for data reconstruction jobs
2012 Operations 10
“Attached” Tier 2 Sites• Page providing updated status of “attached T2 sites” and their storage at
http://lhcbproject.web.cern.ch/lhcbproject/Reprocessing/sites.html– Useful for Tier2 sites to know from/to where they receive/provide the data for
reprocessing– Since this year we have the possibility (and used it) to re-attach sites to another
storage element when processing power was needed
26 Nov '12
2012 Operations 12
User Activities
26 Nov '12
Constant “background” of failed jobs
Higher activities until summer, since then less amount of running jobs
Less submissions during week-end
Job Characteristics• T0/1 sites• Remote input data access• Duration ??
ICHEP
2012 Operations 13
ISSUES AND WORK IN PROGRESS
26 Nov '12
2012 Operations 14
Issues at Sites
26 Nov '12
7 Nov, power Cut at RAL. The site managed to recovered within 24 hours
Mid Oct, disk server failure at CNAF, storage out for several days. After recovery CNAF allotted the double amount of job slots in order to recover
• Mostly “business as usual”– Pilots aborted, memory
consumption by jobs, …
Jobs cannot find their input data, mostly at IN2P3 and GRIDKA, i.e. sites with several “attached” T2s, overload of SRM
2012 Operations 15
Queue Info
• Page providing information on queues as seen from LHCb via BDII– Some sites seem to provide wrong information– As a consequence LHCb is submitting jobs to queues which it shouldn’t
and the local batch system will kill those jobs after max CPU time used• We had to remove temporarily some sites for reprocessing b/c of this
– Currently campaign to cleanup these wrong values ongoing26 Nov '12
http://lhcbproject.web.cern.ch/lhcbproject/Operations/queues.html
2012 Operations 16
Interaction with Sites
• Several sites inform about major downtimes well in advance– Very welcome as it facilitates mid-term planning
• How to reach Tier2 Sites?– LHCb does not have infrastructure to interact
constantly with Tier2s– Can we involve some (wo)men in the middle for
this interaction?• Eg. provide info on processing plans, BDII issue, …
26 Nov '12
2012 Operations 17
CVMFS Deployment• CVMFS deployment is high priority for LHCb (+WLCG)
– Once we reach 100 %will facilitate the sw distribution process
– All sites supporting LHCb are highly encouraged to install CVMFS
– Currently 45 out 96 siteshave CVMFS deployed
– Status info available athttps://maps.google.com/maps?q=http://cern.ch/lhcbproject/CVMFS-map/cvmfs-lhcb.kml
26 Nov '12
2012 Operations 18
OUTLOOK
26 Nov '12
2012 Operations 19
Next Processing Activities
26 Nov '12
Activity Approx Time + Duration
2012 data reprocessing Sep ‘12 – Jan ’13
2011 data reprocessing Beginning ’13 (~ 1 ½ months)
Incremental stripping ~ 2 x / year in 2013 (~ 2 months)
2011/12 data reprocessing During 2014 (~ 5 months)
Loads on sites storage systems• Reprocessing: Reconstruction + Stripping + Merging
– Reconstruction run on “attached T2 sites”– Staging all RAW data from tape – Reco output (FULL.DST) migrated to tape (via disk BUFFER)– Replication of Merging output (DST) on multiple sites
• Incremental Stripping: Stripping + Merging– Staging of all FULL.DST files– Producing up to ~ 20% additional DST files– Replication of DST on multiple sites
2012 Operations 20
Conclusions
• Very good support by sites for LHCb operations– Very good interaction with Tier 1 sites– Improvements possible for Tier2s
• LHCb has made good use of the provided resources in 2012– Upcoming reviews of computing model and tools will have
impact on processes next year • 2012 Reprocessing was good stress test for future
operations– Changes in site infrastructures necessary for post LS1
26 Nov '12
2012 Operations 21
BACKUP
26 Nov '12
FULL.DST = reconstructed physics quantitiesUNM.DST = temporary output for physics stream, PHY.DST = file ready for physics user analysis
Data Processing Workflow
Reconstruction
Stripping
Merging
FULL.DSTUNM.DST
PHY.DST
RAWRAWFULL.DST FULL.DST
Replication
Destroy
TapeTapeBufferDisk Only
Buffer
D1T0
D0T1
Data File
Data Processing
Data Managemt
Disk OnlyStorage
1
2
3
4
5
4’4’
4’6
7
8
10
9
2012 Operations 23
Additional Info• Dirac Job Status plots by final minor status don’t include the Statuses
“Input Data Resolution” and “Pending Requests” because these are not final statuses
• Yandex is not included in the CPU pie plots because it provides sometimes wrong (too high) info on running jobs and would dominate all plots
26 Nov '12