GRID-enabling BEAMnrc & 1 st CLASS PARTICLE TRANSPORT

Transcription

1 GRID-enabling BEAMnrc & 1 st CLASS PARTICLE TRANSPORT mary.chin@physics.org PW CHIN, DG LEWIS & J GIDDY In collaboration with

2

3 RESEARCH THE VISION MONTE CARLO CLINIC

4 MONTHS PER CASE THE VISION DAYS PER CASE MONTE CARLO RUNTIME HOURS PER CASE ~ CALCULATOR

5 WAYS OF MAKING IT FAST! APPROX TRANSPORT e.g. variance reduction cut corners during simulation DENOISING massage data after simulation PARALLEL COMPUTING

6 LESS CONTROL MORE RESOURCES HETEROGENEOUS DECENTRALISED LESS SECURE GRID PARALLEL COMPUTING DISTRIBUTED CLUSTER SINGLE-HOST

7 THE GRID IDEALLY LIKE POWER GRIDs, GENERATION & SUPPLY SHOULD BE HIDDEN FROM THE USER UK e-science GRID single command

8 THE GRID HOWEVER INTERFACING IS COMPLEX & APPLICATION- DEPENDENT UK e-science GRID single command

9 COMBINED MHz CLUSTER GRID PARALLEL COMPUTING SINGLE-HOST # OF PARALLEL RUNS

10 HOWEVER HUMAN LABOUR CLUSTER GRID PARALLEL COMPUTING SINGLE-HOST # OF PARALLEL RUNS

11 SOME LOGIN TO JOBS EACH SITE? IF SINGLE WOULD 10 SITES, THEN GET 10 USERNAMES COMPLETE COMMAND-LINE STUCK & 10 PASSWDS! GET 100 NO. OF FROM THE USER. OUTPUT FILES MESS! NO FURTHER SEND JOBS TO 200 INTERACTION OF WHICH SOME DIFFERENT WOULD BE SITES: CARDIFF, BIG-ENDIAN, GLASGOW,. OTHERS SMALL NEEDED. eg. 200 PARALLEL RUNS SOLUTION

12 LOGIN TO EACH GLOBUS SITE? IF 10 SITES, THEN 10 USERNAMES & 10 PASSWDS! SEND JOBS TO 200 DIFFERENT SITES: CARDIFF, GLASGOW,. SOME JOBS WOULD GET STUCK CONDOR PERL UTILITY TOOLS SOLUTION GET 100 NO. OF OUTPUT FILES OF WHICH SOME WOULD BE BIG-ENDIAN, OTHERS SMALL

13 SOURCES GLOBUS authentication, authorisation, data transfer CONDOR queuing, scheduling, priority scheme, resource classification PERL UTILITY TOOLS streamlining BEAMnrc simulations Chin PW, Lewis DG and Giddy J Implementation of BEAMnrc Monte Carlo simulations on the GRID 14th Int. Conf. on the Use of Computers in Radiation Therapy 2004

14 AUTOMATED

15 Event-by-event ANALOG simulation of each physical process SIMULATION RADIATION TRANSPORT CALCULATION IDEAL BUT NOT PRACTICAL e.g. in gold, electrons undergo 7000 elastic scatterings from 500 to 250keV DETERMINISTIC EQUATIONS

16 ANALOG SIMULATION RADIATION TRANSPORT CALCULATION GENERAL-PURPOSE MONTE CARLO CODES VARIANCE REDUCTION MORE ACCURATE BUT SLOWER DEFAULT QUICKER BUT LESS ACCURATE DETERMINISTIC EQUATIONS

17 an example 1 ST CLASS TRANSPORT USING BEAMnrc 2000 HISTORIES NEED 2x10 9 HISTORIES REQUIRES ~ 6 MONTHS OF RUNTIME ON A 2.8GHz Pentium 4! SCORES 1 PARTICLE WANTED: 10 6 PARTICLES ~ 1 to 2 WEEKS ON THE WELSH e-science GRID (20~60 SGIs) depending on availability of resources

18 1 ST CLASS DEFAULT

19 IN-PHANTOM no noticeable difference, but WATER PHANTOM ISN T THE MOST SENSITIVE TEST

20 PHASE-SPACE ANALYSIS: ENERGY FLUENCE ~ 4% difference

21 PHASE-SPACE ANALYSIS: ENERGY SPECTRUM ~10% peak

22 INVESTIGATION IN PROGRESS: WHO S THE CULPRIT?

23 Summary GRID-enabled BEAMnrc & DOSXYZnrc developed Perl utilities suite for singlecommand automation demonstrated significant difference between 1 st class and default transport parameters Further work Grid of grids! (Presently Welsh e-science GRID only) Acknowledgement Thanking Cancer Research Wales & Yr Ysgol Uwchradd Tregaron for PhD support.