a complete weather forecasting suite on param 10000

21
A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000 Presented by- Sukanta pal Rgdn. No-0701106280 branch-IT 1

Upload: sukanta-pal

Post on 26-Dec-2014

80 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

A COMPLETE WEATHER FORECASTING SUITE ON PARAM

10000

Presented by- Sukanta pal Rgdn. No-0701106280 branch-IT

1

Page 2: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Agenda Introduction. Need for Weather Forecasting. PARAM 10000. Design Approach. Numerical algorithm. The Weather Forecasting Capabilities. Accuracy Issue And Fault Tolerance Conclusion. References

2

Page 3: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

INTRODUCTION What is weather forecasting?

Prediction of atmospheric state for a given time and location

Why to forecast? Air traffic Agriculture Marine Private sector Forestry Military application

3

Page 4: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Methods of weather forecast SynopticStatisticalNumerical

4

Page 5: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Numerical Forecasting Computer simulation of atmospheric condition. It includes

i. Data collectionii.  Data assimilation and analysisiii.  Numerical weather predictioniv.  Model output post processing

5

Page 6: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

PARAM 10000

PARAM is a series of supercomputers developed by CDAC. It is India’s first Teraflop Supercomputer.  The major applications of PARAM 10000 includes

1. long-range weather forecasting

2. remote sensing

3. drug design and

4. molecular modelling.

6

Page 7: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Design approach

7

Page 8: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Based on cluster models of SMPs Has 1gb of RAM per node.

Message passing communication software like MPI and PVM.

Low latency, high bandwidth supported communication software

KSHIPRA.

Lightweight communication protocols on the PARAMNET

and MYRINET

8

Page 9: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

9

Page 10: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Numerical Algorithm Derived from conservation laws of mass, momentum, and energy. Two conditions .

If it is a linear quantity, then the representations are straight forward.

If it is a non-linear term, spectral transform methods have to be used.

The computations for spectral algorithms consists of following steps

10

Page 11: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

jm

nk

M

mn

m

nkj

m puu zZ

||

,

limM

Mnkjkjl ezuzu

,,,

11

Page 12: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

limk

l kjlkjm ezu

kzu

1

02

2 ,,1

,

jm

nkjmk

j jkmn pzuzu ,1

1

12

Page 13: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Weather Forecasting capabilities1. Parallelisation Strategy

Involves of Identification of code and subsequent compute intensive parts and their data dependencies

Latitude is taken as the most obvious and clear independent data for computation work in this algorithm.

Major computation is done through the use of Fourier and Legendre transform.

13

Page 14: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

A pair of latitudes is placed on one processor

Each processor computes Fourier coefficients using the inverse Legendre transform and the gaussian grid point values using the inverse Fourier transform

Partial sum of the Legendre transform is performed on each

processor for given latitudes.

Then these sums are circulated among all the processors to have global sum on each of the processor

14

Page 15: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

2 Implementation

CRAY assembly language was replaced by the Temperton's Fortran algorithm.

The important components of this code are the I. I/O, the forecast loop(GLOOPA), the radiation loop (GLOOPR) which is

called once in twelve hours.

II. Physics loop (GLOOPB), FFT and Legendre transform. GLOOPA and GLOOPB .

The data distribution of parallel T80 is along the latitudes distributed over the processors. This is a master-workers model

15

Page 16: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

16

Page 17: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Accuracy Issue Performance strongly linked to the architectural features like clock

speed, main memory bandwidth.

Message sending proportional to message length but independent of node location and network traffic.

Four cpus in a single node responsible for low latency of message passing between the processors.

17

Page 18: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Fault Tolerance Independent Processing can be run independentently using an

interface.

Fault tolerance save re-running of the codes from system failures.

System failure recovered by restarting JCL.

18

Page 19: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Conclusion Parallel Climate models T2l/T42 and Forecast models T80/T126

have been developed.

Achieve fault tolerance and accuracy.

Access to greater number of cpus thereby enhancing scalability.

Give better performances than before.

19

Page 20: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

Referenceshttp://www.iitr.ac.in/centers/CC/pages/

Resources+Param_1000.htmlhttp://www.cdac.in/html/press/3q09/

spot687.aspxhttp://pune.cdac.in/html/secg/cas.aspx http://pune.cdac.in/html/npsf/npsfidx.aspx http://pune.cdac.in/html/npsf/param10k/

benchmp.aspx http://www.buyya.com/papers/cdacrep.txt http://en.wikipedia.org/wiki/

Remote_direct_memory_access 20

Page 21: A COMPLETE WEATHER FORECASTING SUITE ON PARAM 10000

THANK YOU

21