1. ALICE Computing : 2012 operation & future plans
Rencontre LCG-France, SUBATECH Nantes
18-20 September 2012

2. A quick glimpse of 2012 Standard data taking year for ALICE
p-p – emphasis on rare triggers, high Pt (Calorimeter)
pilot p-A run (few million events)
long p-A run in February 2013 (still counts as ‘2012’)
Bulk of analysis on 2011 Pb-Pb - the largest single period data sample

3. 2012 - RAW Standard treatment – 2 copies of RAW data
One at T0, one replica at T1s, proportional to the fraction of mass storage capacity
~1PB of RAW until now

4. 2012 – job profile Average 28K jobs in parallel
Increases as capacities become available

5. 2012 – site contribution
Wall time – 50/50 T0/1 to T2s

6. 2012 – French sites
Wall time – 25/75 T1 to T2s

7. Other important parameters
Storage (always insufficient…)
2PB of disk, 45% at T1
The balance is not as equal as CPU
Network – extremely well provisioned
T2 connectivity
will further
improve with
LHCONE

8. More details on workload
Organized
activities, including
trains
Chaotic

9. Resources use - tasks
 Last year goal – increase the fraction of organized analysis
Tool – analysis trains
Long-term goal, takes a substantial amount of coordination and user education
The resources use distribution
10% RAW reconstruction (constant)
16% train analysis (5% beginning of year)
23% chaotic analysis (36% beginning of 2012)
51% Monte-Carlo productions (49% beginning of 2012)

10. The Analysis Trains
Polling together many user analysis tasks (wagons) in a single set of Grid jobs (the train)
Managed through a web interface by a Physics Working Group conductor (ALICE has 8 PWGs)
Provides a configuration and test platform (functionality, memory, efficiency) and a submission/monitoring interface
Speed – few days to go through a complete period (PBs of data!)
MonALISA
Web interface
LPM
AliROOT
Analysis
Framework
AliEn
Grid jobs

11. Data access in analysis
The chaotic and to some extent organized analysis is I/O bound (efficient use of disk/network resources)
Average 8 GB/s, peak 20 GB/s
Total read data from 1-st of April is 120 PB.

12. CPU efficiency
Stable (but ‘low’), some improvement with time – increase of trains share over chaotic analysis

13. CPU efficiency for organized tasks
MC - high, RAW ~OK, trains – still needs improvement

14. Analysis efficiency Processing phases per event
Reading event data from disk
– sequential
De-serializing the event object
hierarchy – sequential
Processing the event
parallelizable
Cleaning the event structures
- sequential
Writing the output – sequential
but parallelizable
Merging the outputs – sequential but parallelizable
tread
tds
tproc
tcl
twrite
Event #0
Event #1
Event #2
Event #m
Event #0
Event #1
Event #2
Event #n
Event #0
Event #1
Event #2
Event #p
tmerge
A.Gheata – improving analysis efficiency

15. Analysis efficiency (2)
The efficiency of the analysis job:
job_eff = (tds+tproc+tcl)/ttotal
analysis_eff = tproc / ttotal
Time/event for different phases depending on many factors
Tread~ IOPS*event_size/read_throughput – to be minimized
Minimize event size, keep under control read throughput
Tds+Tcl~ event_size*n_branches – to be minimized
Minimize event size and complexity
Tproc = ∑wagonsTi – to be maximized
Maximize number of wagons and useful processing
Twrite = output_size/write_throughput – to be minimized
A.Gheata – improving analysis efficiency

16. Grid upgrades New AliEn version (v.2-20) – ready for deployment
Lighter catalogue structure
Presently M LFNs, 2.5x PFNs (replicas)
Growing at 10Mio new entries per week
Extreme job brokering
The jobs are no longer pre-split and with pre-determined input data set
Potentially one job could process all input data (of the set) at a given site
The data locality principle remains (for now)
The site/central services upgrade – need some downtime, after end of data taking in Feb.2013

17. analyze as much as possible
File Brokering
Site A
Site B
Site C
File 1
File 2
File 3
File 4
File 5
Current schema
Submit 4 jobs:
File1
File 4
File2
File3
File 5
Broker per file
Submit 3 empty subjobs
If nothing left,
just exit
File
1,2,4,5
When a job starts,
analyze as much as possible
File 3
From P.Saiz – AliEn development

18. Short development roadmap
Data management:
Popularity service
SE layout (EOS-like)
GUIDless catalogue
Job Processing:
Job Merging
Error classification
Multicore and multiagent
Remote access optimization
Combine AF/classical grid CE – interactive Grid

19. General remarks on the future
– Long Shutdown 1
No revolution is (ever) planned, however…
All LHC experiments have submitted LoI for the LS3 (HL-LHC) upgrades in 2022
For the computing is massively larger than today (data rates and volumes, CPU needs) – 10-30x of today – the factors are not yet finalized
Massive online DAQ and HLT event filtering farms 2x size of what a T1 is today
No clear ideas how this will be achieved – technologically and financially
Moore’s and Kryder’s laws will not ‘cover’ the needs

20. General remarks on the future
The present Grid profited from ~10 years of planning and development (on par with the detectors)
And it delivered from day 1, continues to this day
The future planning and development of Grid/Cloud/<Insert name here> should start now – years of experience will help, but not enough
Parallel programming cannot be done by physicists… there are other hurdles too

21. General remarks on the future
Big improvement is expected from the frameworks and code
Undoubtedly a common effort and professional help will be necessary
Parallelism is a no-brainer, given the technological trends
Big parts of the code must be re-engineered and re-written
Every experiment has a panel which is charged with the design of the ‘new’ software
Crystal balls have been ordered 

22. Summary – back to today
 The 2012 is so far a standard data taking/processing/analysis year for ALICE – much excitement is expected in February with the p-A data
The operation is smooth and is helped a lot by the mature Grid around the world
The French T1/T2s are part of this structure, with remarkably stable performance and well balanced components – CPU, storage, networks
… and of course a solid expert support at all levels – a big **thank you** for this!

23. Summary – cont.
 The (near)future developments are focused on analysis tasks and tools
Emphasis on data containers and process synchronization
Whole node is a promising path and will naturally help the multicore development
Progressive introduction of new features, improvements
The Grid must run continuously also during the LS1 shutdown
Resources (disk) are scarce
More efficient use – less replicas, WAN access