Cluster Programming Technology and its Application

in Meteorology

Computer and AutomationResearch InstituteHungarian Academy of Sciences

HungarianMeteorologicalService

Silicon Computers Ltd.

Background• Hungarian Meteorological Service developed

MEANDER (MEsoscale Analysis Nowcasting and DEcision Routines)– crucial task in the protection of life and property (storm warning at

Lake Balaton, weather warnings for aviation...)– based on incoming meteorology information and computational

intensive methods

• MTA SZTAKI developed P-GRADE parallel programming environment – efficient, graphical support for the entire life cycle of parallel program

development

• Cluster programming technology and its applications in meteorology– joint project of MTA SZTAKI, Hungarian Meteorological Service and

Silicon Computers Ltd.– supported by Research & Development Division, Ministry of

Education

Goal: Analysis of all the available meteorology information producing parameters on a high resolution regular mesh

(10km--> 1km) ultra-short range forecast (up to 6 hours)

Application: Forecasting dangerous weather situations (storms, fog, etc.)

Meteorology information: surface level measurements, high-altitude measurements, radar, satellite, lightning, results of previous computed models, etc.

Basic parameters: pressure, temperature, humidity, wind, …

Derived parameters: type of clouds, visibility, …

MEANDER Program Package

Structure of MEANDERFirst guess dataALADDIN

SYNOP data Satellite

dataRadardata

CANARI DELTA analysi

sBasic fields: pressure, temperature, humidity, wind

Derived fields: Type of clouds, visibility, etc.

BASIC GRID

Radar to grid

Satellite to grid

“Present” weather

Type of clouds

Overcast

Visibility

Rainfall phase

Visualization For meteorologist:HAWKFor users:

GIF

Lightning

decode

36000 km

Satellite raw image

Receivingimage

Transformation & Interpolation &

Processing

Processing:altitude of clouds Transformation Interpolation to basic GRID

Processing of satellite images

THE PROBLEM

MEANDER Sequential

code

C, C++ Fortran

MEANDERParallelversion

PC clasterSGI Origin 2000

SUN E10000

Parallelisation?

Debugging?

Performance?

GRAPNELGRAPNEL graphical language / GRP2CGRP2C pre-compiler

Unix-like operating systems & C, C++, Fortran programming libraries

PVM or MPI message passing library

DIWIDEDIWIDE

distributed debugger

GREDGREDgraphical

editor

GRMGRM monitor

ddistributedistributed ccheckpointingheckpointing

tooltool

GRP_CHKPTGRP_CHKPT

GRP_MMGRP_MM mmigration moduleigration module

GRP_LBGRP_LB lload oad bbalanceralancer PROVEPROVEperformancevisualization

tool

TLCTLC modelchecker

MacrostepMacrostepdebugger

The solution: P-GRADE P-GRADE development

environment

Results: temperature and wind at 850 hPa level

Results: temperature and wind at 850 hPa level

MEANDER: 3D FIELDS

2D analysis

Radardata

Satellitedata

Delta 3Danalysis

...computes the basic meteorological fields:

pressure, temperature, humidity, wind velocity and direction

for a high resolution 3D

mesh (10km -1km)

Implementation of

DELTA analysis in

P-GRADE

Fortranseq. code

P-GRADE version of MEANDER for clusters & supercomputers

25 x

4 x

10 x

25 x

25 x

20 x

Live demo (5th DataGRID conference)

34 MbitShared

PERL-GRIDCONDOR-PVM

job

11/5 MbitDedicated

P-GRADEPERL-GRID

jobftp.met.hu

netCDF

512 kbitShared

netCDFinput

Parallel execution and GRM

GRMTRACE &Results

GRMTRACE &Results

PROVE performance visualization

Edit, debugging

Performance-analysis

Testing,Execution

Resource requirements of P-GRADE

Advantages of P-GRADE environment Efficient support for each stage of parallel

program development Fast parallelisation of existing algorithms

Reusability of sequential code Hiding of low level communication functions Unified and integrated graphical concept Predefined communication templates Support for hierarchical design

Even non-professional programmers can use it

(steep learning curve) Portability from supercomputers to PC clusters

Support for cluster computing

More information: www.lpds.sztaki.hu

New facilities in P-GRADEfor long-running parallelapplications:

• Distributed checkpointing• Process migration• Dynamic load balancing• Fault-tolerance execution

I. Execution withoutload balancing

III. Load balancing& migration

IV. Result:Faster execution

II. Checkpointing