1. A Grid fusion code for the Drift Kinetic Equation solver A.J. Rubio-Montero, E. Montes, M.Rodríguez, F.Castejón, R.Mayo CIEMAT. Avda Complutense, 22. Madrid (SPAIN) Abstract Several Drift Kinetic Equation solvers (DKEs) codes are used to simulate the neoclassical transport for the stellerators and tokamaks fusion reactors. These applications offer an approach more near to analytic solution than Monte Carlo methods, but usually requiring a great amount of memory, making their porting to Grid difficult. This work explains the gridification process of a DKEs code widely accepted by the Fusion community. Also, performance and portability gains are evaluated. The tests and results obtained have been applied to the TJ-II Flexible Heliac at National Fusion Laboratory (Spain) by using mainly the EELA-2 infrastructure. Impact, objectives and Grid-added value Calculating neoclassical transport Is a fundamental part of the complete simulation cycle of the behaviour of plasmas inside fusion reactors. It is also usually used to determine the energy flux to the plasma wall for a certain magnetic field, and therefore, the efficiency of a certain coil configuration before it is implemented (i. e. ITER). Neoclassical transport can be calculated by means of Monte Carlo methods, which have been successfully deployed on Grid to solve a wide range of scientific challenges in many disciplines, among them Physics, but for this specific case they only can offer an estimation of the perpendicular diffusive part of the transport matrix. On the other hand, DKE solvers provide correct quantitative results of the complete transport matrix, with the drawback of high computation time and memory consumption, being usually executed on shared memory computers. Then, the objective is offer to Scientifics a new framework, called DKEsG (Drift Kinetic Equation solver for Grids), to easily retrieve all neoclassical transport coefficients:  Reducing time execution and increasing precision making use of heterogeneous resources from different Grid infrastructures.  Filling a database with the complete “configuration - transport matrix – state” for several fusion reactors to:  Bring results to Fusion Community  Avoid performing the same simulation twice and easy re-analysis  Easy interaction with other fusion applications to build complex workflows Original Variational DKES code [1]  Applied to TJ-II, HSX, CHS, LHD, ATF, VII-AS, VII-X, NCSX, QPS, ITER and other devices.  Implemented for obsolete shared memory hosts (CRAY-1, CRAY-X-MP) and modified to run on SGI Origin 3000 MIPS based machines.  Requires continuously manual configuring, recompiling code and compilation of high number of output data files  Calculates only diffusion coefficients for a certain number of collisionality/energy and radial electric field values Current Status [2]  All code in a nutshell  Algorithms from DKES software up-to date and extended  Ported and optimized to Linux x86-32, x86-64 and ia-64 platforms  Only a binary per architecture and algorithm  No software must be installed in resources  Grid-enabled: Run on Globus and gLite based Grids, being tested on EGEE-III, EELA-2 and regional infrastructures. Future  High level implementation on java DRMAA standard  Suitable to run on Globus/gLite Grids through GridWay or on local clusters through PBS, SGE or LSF  Integration in complex application level workflows and databases  Independent GUI  A new module to calculate transport coefficients has been developed  Fill coefficients database and offer it as virtual observatory References [1]W.I. van Rij and S.P. Hirshman. “Variational bounds for transport coefficients in three-dimensional toroidal plasmas”, Phys. Fluids B 1 (3), 563-569 (1989) [2]A.J. Rubio-Montero, E. Montes, F. Castejón, L.A. Flores, M. Rodríguez, R.Mayo. Calculations of Neoclassical Transport on the Grid. Proceedings of First EELA-2 Conference, Vol.1, p.120-125. 2009. Some tests calculating diffusion coefficients  Task description  5 collisionallity variations per task  Fixed electric field  100 Legendre and 343 Fourier polynomials  ~450 seconds on a Xeon(2006) 3.2 GHz  ~285 seconds on a Itanium2(2005) 1.5GHz  1ms of simulation requires more than 10 6 jobs ( > 14 years in a single Xeon processor)  10 3 jobs spent experimentally 4h 20’ on a heterogeneous Grid with only 60 available processors, ~34% less than 50 processors from SGI Origin 3800. Neoclassical transport coefficients for all possible magnetic surface, temperature, density are indexed in a public database by fusion device. Diffusion coefficients calculated by DKES by collisionallity for a single electric field SGI 3800 (50 MPI processors 600Mhz) CIEMAT Internal Grid (limited to 25 slots among 5 heterogeneous resources) EELA + Internal + others (limited to 60 slots among 12 available resources) Jobs completed Performance (jobs x second) 36.8% 105.2%