CompHEP development for distributed calculation of particle processes at LHC

1
CompHEP development for distributed calculation
of particle processes at LHC
1

• A.Kryukov(kryukov_at_theory.sinp.msu.ru),
  L.Shamardin(shamardin_at_theory.sinp.msu.ru)Skobelts
  yn Institute of Nuclear Physics,Moscow State
  University

2
Outlook
2

• Motivation
• Modification of CompHEP for Distributive
  Computation
• Conclusions and
• Future Development

3
Motivation
3

• Specific of hadron collider physics is a huge
  number of diagrams, hundreds of subprocesses.
• SM (Feynman gauge) p,p -gt W,W-, 2jets
• pu,U,d,D,c,C,s,S,b,B,G
• jetu,U,d,D,c,C,s,S,b,B,G
• QCD background (excluded diagrams with virtual
  A,Z,W,W-)
• Number of subprocesses is 775
• Total number of diagrams is 16461

4
Motivation (cont.)
4

• Possible solution is simplification of the model
  combinatory Boos, Ilyin.
• u-d-model (Feynman gauge) p,p -gt W,W-, 2jets
• pu,U,d,D,b,B,G
• jetu,U,d,D,b,B,G
• QCD background (excluded diagrams with virtual
  A,Z,W,W-)
• Number of subprocesses is 69
• Total number of diagrams is 845

5
u,U -gt d,D,W,W-
5
6
CompHEP calculation scheme
6
Diagram generator
C-code generator
Compilation
Symbolic calculation

• Generate single executable file for all sub
  processes
• Nothing utilities to make easy tuning of MC
  integration for different sub processes.

Sub process selection
Next sub process
MC calculation
7
Modified Scheme of calculation
7
C-code generator
Diagram generator
Symbolic calculation
...
...

• Separate executable file for each sub process.

8
File structure
8

• Old file structure was

Modified file structure
Working Dir.
Working Dir.
Results
Results
F1.c
Sub1
F2.c
F1.c
...
F2.c
...
...
Sub2
F43.c
F44.c
...
...
...
9
Results
9

• Hardware P4 (1.3GHz), RAM256M
• Process p,p -gt W,W-,2jet
• 69 sub processes, 845 diagrams
• ud model
• Standard Distributed, Distributed,
• (Per sub proc.) mean value total
• Size of executable file71M (1.02M) 2.9M 200.1M
• Compilation time 176m (153s) 133s 53m

10
Results (cont)
10
Standard Distributed Stnd/Dis. (Per sub
proc.) (mean value) Cross section calc. 22m
(19s) 15s 1.3 Memory (Virt./RAM) 46.5M/1.8M 7.0M
/2.0M 6.7/0.9 Maximum search 60m(52s) 50s 1.1
Memory (Virt./RAM) 46.5M/1.8M 6.8M/1.8M 6.7/1.0
Generation of 1kev. 106m(92s) 60s 1.5 Memory
(Virt./RAM) 46.7M/2.1M 7.0M/2.0M 6.7/1.1
11
Specific features to support distributed
calculation
11

• Copy session data of selected sub process to all
  other sub processes.
• Copy individual parameter of selected sub process
  to all other sub processes.
• Copy session data from specific sub process to
  selected sub process.
• Copy specific parameterfrom selected subprocess.

12
Specific features to support distributed
calculation (cont.)
12

• Utility for job submission under PBS and GRID
  (LCG-1).
• Modified utility that collect event stream
  generated by separate subprocess executables into
  single event sample.
• d_comphep -cCPG path_to_results_dir
• -c compilation only
• -C collect data into single sample
• -P PBS (default)
• -G GRID

13
Conclusions and ...
13

• Hadron collider physics require to use computer
  cluster and/or GRID solution for calculation with
  more then 2 hard interacted particle in finale
  state.
• It is necessary to develop special tools to
  support such kind calculations.
• Even rather simple 2-gt4 process has profit if we
  use distributed computation. Here we do not
  discuss the problem of convenience for user.

14
Future Development
14

• In this work we realize rather straightforward
  approach of distributive computation in CompHEP.
  In the future we are going to consider more
  sophisticated method takes into account the phase
  space as a set of non-overlapping pieces. This
  approach permit to divide any task on those
  number of independent (from MC point of view) sub
  tasks as necessary.