Presentacin de PowerPoint

1
Interactive High Energy Physics Application
Celso Martinez-Rivero Jesus Marco, Rafael
Marco, Oscar Ponce David Rodriguez IFCA-CSIC
(SPAIN) CROSSGRID WP1 MEETING 18-III-2002 KRAKOW
2
Distributed Physics Analysis in HEP

• Final user applications for interactive physics
  analysis in a GRID aware environment for LHC
  experiments
• Access to large O/R DBMS through middleware
  server vs. access to catalogue distributed root
  data files
• Distributed data-mining techniques (mainly
  Neural Networks, also self organizing maps, ...)
• Integration of user-friendly interactive access
  via portals
• CSIC UAB FZK INP INS

3
HEP interactive analysis

• USER REQUIREMENTS
• LHC experiments
• Other taking data? CDF, BaBar
• Physics analysis physics results (TDR,
  publications)
• HLT Trigger optimisation
• TASKS
• 1.3.1 Interactive Distributed Data Access
• 1.3.2 Data Mining Techniques
• 1.3.3 Integration and Deployment
• 1.3.4 Application to LHC Physics TDR
• DEVELOPPERS
• CSIC D.Rodriguez (O/R DBMS),O.Ponce,
  C.Martinez,J.Marco
• FZK M.Kunze people on ROOT side
• CMSCSIC (Santander) C.Martinez
• ALICEFZK M.Kunze
• ATLAS UAB A.Pacheco, M.Bosman CSIC (Valencia)
  E.Ros, J.Salt, INP P.Malecki
• LHCB INP M.Witek

4
The final product GUI
Monitoring
Graphic Output/(Input?)
DATASET Dictionary (Classes) Basic
Object Derived Procedures
Alphanumeric Output
Analysis Scripts
Work Persistency
5
(XML) Info Work flow
UI
Replica/Data Manager Web Service Finder?
Broker
Master CE
CACHED
6
Storage Element as WebService?

• Current SE in EDG
• GridFTP server
• WebService approach
• Passive SE GridFTP, or /grid , etc...
• Active SE
• SQL QUERY (ResultSet in XML) SELECT FROM
• (Three tier servlet running, like Spitfire)
  ready! (IBM IDS)
• ROOT query (does this make sense? Paw query
  does make sense, implemented...)
• PROCESSING QUERY ( Agent) Stored Procedure or
  XML description (SOAP like?)
• SQL QUERY ok for NN in HEP
• PROCESSING QUERY (Agent-like approach) needed
  likely for SOM

7
Data Mining

• Basic case 1
• Neural Network, BFGS
• Objective Parallel Training
• Balance Data Load (How? Split AFTER DB Query)
• MPI works ok for distributed calculus !
• How to do in ROOT environment (PROOF?)
• First Results...
• Basic case 2
• SOM (Unsupervised learning)
• Objective Cluster Analysis
• Wait for Meteo 1.4.b experience

8
Parallel NN

• Prototype
• MLP package in PAW (databasen-tuples)
• Used for DELPHI Higgs search (REALISTIC)
• BFGS method
• Current setup
• Scan ntuples to filter events variables
• Result set in XML, split according to CEs
• 1 Master n slaves, in LOCAL cluster, using
  MPICH G2
• NN architecture 16-10-10-1
• Master sends initial weights, each node returns
  gradient and errors to master who adds and sends
  new weights...
• Scaling ok in local cluster, now moving to
  non-local environment with latency and QoS in
  mind!
• Obvious solution adapt NN load in each node to
  latency time

9
NN scaling
Lattency curve
644577 events, 16 variables 16-10-10-1
architecture 1000 epochs for training
10
Integration

• Final User applications should run on simulated
  data from the 4 LHC experiments.
• TBD

11
Example of Trigger Analysis in CMS

• HLT Trigger almost equivalent to offline physics
  analysis
• 3 Trigger levels under study
• L1 Hardware
• L2 Software standalone
• L3 Software full reconstruction

12
Example Single m stream
13
Fine Tuning of HLT using Crossgrid

• Some basic topologies with no need of high level
  analysis (isolated high Pt muons)
• However other topologies much more difficult to
  separate from background
• More sophisticated analysis mandatory to keep
  background rate low while high efficiency!
• Crossgrid will help here!

14
Trigger CPU Time

• L2 780 ms average
• Main processgt Trajectory Builder
• Important queues!
• L3 1689 ms average large fluctuations mostly
  combinatorial.
• Reduce muon cone?

15
Trigger CPU Time

• What to do with large CPU consuming events?
• Use Grid to avoid collapsing the standard
  trigger farms.
• All events requiring more time to process than a
  certain limit time are automatically sent to
  Crosssgrid for triggering

16
What do we need from Tools Services

• GRID Application Programming Environment
• Verification of MPI use YES
• Performance prediction YES
• Monitoring YES (action if a node is down)
• Services Tools
• User friendly portals YES (we go for all XML)
• Roaming access YES (but implicit in portal)
• Efficient distributed data access YES (need close
  contact)
• Specific resource management YES (general
  question, how to manage interactive parallel
  jobs)

17
Short Answer to WP2 questions

• Programming languages
• C, C,Java
• HLA no
• CCA could be, not priority
• Component structure
• Matrix involved in gradient for NN
• Granularity
• Important if NN becomes too big, and in SOMs
• Performance problems
• Main Latency
• Monitoring
• Check dead nodes, lets think what to do
• Node time for processing, Latency
• Storage element performance

18
Short Answer to WP2 questions

• MPI yes
• Calls (initial list)

Now MPI_Init(argc, argv) MPI_Comm_size(MPI_COM
M_WORLD, nproc) MPI_Comm_rank(MPI_COMM_WORLD,
rank) MPI_Get_processor_name(processor_name,nam
elen) MPI_Barrier(MPI_COMM_WORLD) MPI_Bcast(buff
er_int, 1, MPI_INT, 0, MPI_COMM_WORLD) MPI_Finali
ze() MPI_Recv(buffer_dbl,num_var , MPI_DOUBLE,
i, 30, MPI_COMM_WORLD, status) MPI_Send(buffer_d
bl, num_var, MPI_DOUBLE, 0, 30,
MPI_COMM_WORLD) Soon Next MPI_Reduce(buffer_dbl
_slave, buffer_dbl_master, num_datos, MPI_DOUBLE,
MPI_SUM , 0, MPI_COMM_WORLD)
19
What do we need from Testbed

• Definition of Interactive resources?
• Reserved/Priorized?
• Cached storage(?)
• Submission or Interactive session?
• Check latency and if possible impact of QoS (see
  map)
• Need about 3-5 testbed sites each one with 50
  processors to make a realistic test of the
  potential.

20
CrossGrid WP4 - International Testbed Organisation
Network (Geant) setup