PARALLEL IMPLEMENTATION OF PARTON STRING MODEL EVENT GENERATOR

 A.Asryan, G.Feofilov, S.Nemnyugin, P.Naumenko,

V.Solodkov, V.Vechernin, A.Zarochencev, V.Zolotarev

Saint-Petersburg State University,Russian Federation

Table of contents

Problem (what is the matter?)ModelPSM event generatorWhy parallel?Parallel PSM event generatorPerformance testsResults of simulation

Keywords of the report:

software

Monte-Carlo event generator PSM(Parton String Model)

The method of simulation is based on the Parton String model with strings fusion taken into account.

The modified model has been suggested by M.A.Braun (Saint-Petersburg State University) and C.Pajares (University Santiago de Compostela).

The model was realized as PSM Monte Carlo event generator by N.Amelin (JINR, Dubna).

M.A.Braun, C.Pajares, Phys. Lett. B287 (1992) 154; Nucl. Phys. B390 (1993) 542, 549.

N.S. Amelin, M.A.Braun, C.Pajares, Phys. Lett. B306 (1993) 312; Z.Phys. C63 (1994) 507.

PSM PROGRAM GENERATORSHORT DESCRIPTION: FIRST VERSION IS WRITTEN BY DR. N. AMELIN FROM THE JOINT INSTITUTE FOR

NUCLEAR RESEARCH (DUBNA, RUSSIA) PURPOSE  TO SIMULATE NUCLEON-NUCLEON, NUCLEON-NUCLEUS AND NUCLEUS-

NUCLEUS COLLISIONS AT ULTRA-RELATIVISTIC ENERGIES MODEL  PARTON STRING MODELSee N. Armesto, M. A. Braun, E. G. Ferreiro and C. Pajares, Phys. Lett. B344 (1995) 301 SOME MODEL DETAILS:GLUON RADIATION AND HARD GLUON RESCATTERINGS ARE INCLUDED BEAM AND TARGET MAY BE: N, P, anti-P, D, He, Be, B, C, O, Al, Si, S, Ar, Ca, Cu, Ag,

Xe, W, Au, Pb, U

 RANGE OF ENERGIES (in Gev): 10 < Ecm / NUCLEON < 15 000 IMPACT PARAMETER: FIXED OR RANDOM NUMBER OF EVENTS: > 1 STRING FUSION: MAY BE INCLUDED GLUON RESCATTERING: MAY BE INCLUDED BUT IS EXTREMELY TIME CONSUMING IT IS POSSIBLE TO SIMULATE +, –, K+, K– INDUCED REACTIONS METHOD OF SIMULATION: MONTE CARLO METHOD 

PSM PROGRAM GENERATOR

Long-range correlations

Features of installation taken into account

WHY PARALLEL?

A lot of statistics (104 - 106 of simulated events) is required in a Monte Carlo simulation to get statistically reliable results.

WHY PARALLEL?

Simulation is time consuming if all options of the model are turned on:

hard gluon rescattering; string fusion; resonances decay; rescattering of secondaries (sourced from

string breaking) + secondaries and spectators (!).

WHY PARALLEL?

13.5 secs/event on 600 MHz CPU with all options of the model turned on.

Time-consuming? -> supercomputing/parallel programming should be applied.

SIMULATION WITH THE PSM MONTE CARLO GENERATOR

1. Getting “raw” output data with the PSM Monte Carlo generator

2. Processing of the output file by the PERL program (new). The result is a correlation diagram as the Postscript file

Parallel algorithms of PSM modeling have been developed

and tested on parallel clusters of Saint Petersburg State

University.

Parallel version of the PSM Monte-Carlo generator is realized with MPICH

library (version 1.2.4) 

PROBLEMS TO BE SOLVED

1. Effective parallelization of the PSM Monte Carlo generator:configuration of cluster is important

2. Effective output operations:25 000 events ~ 1 Gbyte output file

500 000 events ~ 20 Gbytes output file3. Pseudorandom number generator:    parallel, long enough period, good statistical

properties (low correlations) etc.

PARALLEL PSM GENERATOR FOR CLUSTER WITH LARGE LOCAL

DISKS

Master-slave model of parallel programming.

Distribution of events to be simulated between nodes of the parallel cluster.

Master process coordinates work of other (slave) processes.

Master process broadcasts (scatters) input data.

Results of modeling are saved in local files (/tmp directory of a hosts). 

PARALLEL PSM GENERATOR FOR CLUSTER WITH LARGE LOCAL

DISKS

After completion of the PSM program local files are sent to the “master” host.

The program (Perl) of statistical data processing is started.

Result (correlation diagram) is saved in the file in Postscript-format.

PARALLEL PSM GENERATOR FOR CLUSTER WITH LARGE LOCAL

DISKS

Pro:

Communication network is not overloaded -> maximum scalability, theoretical limit of Ahmdal’s law

Contra:A lot of work when merging local output files

PARALLEL PSM MONTE CARLO GENERATOR FOR CLUSTER WITH FILE

SERVER

Master-slave model of parallel programming.

Parallel output in the shared file. It is located on the file server.

Loading of a communication subsystem of the multiprocessor computer grows.

PARALLEL PSM MONTE CARLO GENERATOR FOR CLUSTER WITH FILE

SERVER

Pro:More pleasant “postmortem” life

Contra:Communication network is overloaded > poor scalability

PARALLEL PSM MONTE CARLO GENERATOR FOR CLUSTER WITH FILE

SERVER

Another approach - compromise:

not use parallel output operations

use gathering operations.

HARDWARE*

“ALICE” cluster in Saint Petersburg:

7 x 2CPU hosts (2х600 MHz, 512 Mbyte RAM, 2х4,5 Gbyte hard disks)

+1 server (1200 MHz, 256 Mbyte RAM, 40 Gbyte hard disk)

---------------------------------------------------* Before September 2004, now 1 Gbyte RAM, 40 Gbyte

hard disks

SPEEDUP TEST (1st approach)

Speedup vs Number of processors for different PSM options (speedup=1 CPU time/N CPUs time) – ABCD:

A) HARD PART: True/False; B) STRING FUSION;

C) RESONANCE DECAY; D) RESCATTERING

PERFORMANCE TESTS1000 eventsTTTT--------------------------------------------------------------- HOSTs time---------------------------------------------------------------

1 224m21.353s 2 114m13.243s 3 87m59.043s 4 62m.83.205s 5 54m67.012s 6 48m71.642s 7 44m32.502s

PERFORMANCE TESTS1000 events TTTF--------------------------------------------------------------- HOSTs time---------------------------------------------------------------

1 14m46.613s

2 7m59.188s

3 5m24.613s

4 4m16.065s

5 3m35.322s

6 3m9.149s

7 2m49.423s

PERFORMANCE TESTS1000 eventsTTFF--------------------------------------------------------------- HOSTs time---------------------------------------------------------------

1 14m32.460s

2 7m24.718s

3 4m47.604s

4 3m41.443s

5 3m11.629s

6 2m41.460s

7 2m13.166s

PERFORMANCE TESTS1000 events -TFFF--------------------------------------------------------------- HOSTs time---------------------------------------------------------------

1 4m35.410s

2 2m20.976s

3 1m35.235s

4 1m10.788s

5 0m58.430s

6 0m46.027s

7 0m41.621s

RESULTS OF SIMULATION

RESULTS OF SIMULATION

PSEUDORANDOM NUMBERS GENERATOR

Copy of the CERN Library routine RECUSQ

Authors: R.Brun, F.Carminati

Up to 215 pseudorandom sequences

Each sequence has a period of 10**9 numbers – good for 104 105 statistics, not so good for 106 statistics.

PSEUDORANDOM NUMBERS GENERATOR

SUBROUTINE RANLUX(RVEC,LENV)C Subtract-and-borrow random number generator proposed byC Marsaglia and Zaman, implemented by F. James with the nameC RCARRY in 1991, and later improved by Martin LuescherC in 1993 to produce "Luxury Pseudorandom Numbers".C Fortran 77 coded by F. James, 1993C references: C M. Luscher, Computer Physics Communications 79 (1994) 100C F. James, Computer Physics Communications 79 (1994) 111C LUXURY LEVELS.C ------ ------ The available luxury levels are:C level 0 (p=24): equivalent to the original RCARRY of Marsaglia and Zaman, very long

period,C but fails many tests.C level 1 (p=48): considerable improvement in quality over level 0, now passes the

gap test,C but still fails spectral test.C level 2 (p=97): passes all known tests, but theoretically still defective.C level 3 (p=223): DEFAULT VALUE. Any theoretically possible correlations have very

smallC chance of being observed.C level 4 (p=389): highest possible luxury, all 24 bits chaotic.

CONCLUSIONS/FURTHER PLANS

Parallel PSM event generator is developed. Parallel PSM event generator works on “ALICE”

computing cluster (Saint Petersburg State University).

It is scalable. Parallelization methods for different cluster

configurations are analyzed. “Compromise” version is under development. High-performance -> more statistically reliable

results + more realistic physical model