OMD 2 platform dedicated to HPC Optimisation

OMD 2 platform dedicated to HPC Optimisation ScilabTEC 2011 École Polytechnique ACKNOWLEDGMENT: This work has been supported by French National Research Agency (ANR) through COSINUS program (project OMD 2 n ANR-08- COSI-007) Guillaume Jacquenot, SIREHNA Contributors : Luc Bordier, SIREHNA Vincent Couvert, DIGITEO SCILAB Maryan Sidorkiewicz, RENAULT (coordinator)

INTRODUCTION Top 500 Ranking of the 500 top supercomputers in the world Deliver trends in high-performance computing 1/29 => Increase of power computation allows to consider new simulation possibilities Source : www.top500.org Classical CPU : 5GFlop/s

SUMMARY 01 02 03 OMD2 PROJECT : HPC & CFD OPTIMISATION INDUSTRIAL CONTEXT, OBJECTIVES FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW 04 OUTLOOK DELIVERABLES, WORK IN PROGRESS, DISSEMINATION 2/29

OMD2 PROJECT : HPC & CFD OPTIMISATION 01 CONTEXT Economical and environmental issues Product development processes are regularly shortened Car manufacturing specific problems: Reduce pollutant emissions and fuel consumption CO2 and EURO VI regulation Numerical optimization issues Computation time Decision making tool Multidisciplinarity 3/29

OMD2 PROJECT : HPC & CFD OPTIMISATION 01 HPC ACTIVITY Main objectives : Reduce mock-up process and improve size and precision of modelling Innovate and improve customer requirements Reduce simulation time Field of applications: Crash, powertrain, CFD : External aerodynamic, Exhaust line, HVAC Source : Renault & Sirehna 4/29

01 OMD2 PROJECT : HPC & CFD OPTIMISATION FOCUS ON CFD APPLICATIONS & OBJECTIVES Some figures Ship engineering : Small variations on a ship hull can increase wave resistance of several percents Car engineering : +/- 1mm of manufacturing dispersion on the cylinderhead => - 20% of performance (worst case) Main difficulties Lack of automation of the numerical process Expensive cost of HPC licenses Mainly local optimization Work in progress Large scale optimization techniques Topological optimization Software benchmarking Cloud computing Source : Renault 5/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION List of test cases from OMD2 project 2D case CFD HVAC pipe 3D case CFD HVAC pipe 3D case CFD Rear-view mirror 3D case CFD External aerodynamics on a car 3D case CFD Cylinderhead 3D case CFD Ship hull Source : Renault & Sirehna 6/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION 2D case HVAC pipe Objective of test case: Test of methods and algorithms proposed : - Hybrid optimisation with use of partial convergence algorithm (Developed by V. Picheny at ECP) - Prepare procedures for 3D complex cases Objective of the problem : Find the optimal geometry to : F1 : Minimise the pressure loss between inlet and outlet F2 : Minimise of the standard deviation of the output speed profile Geometry described with 13 parameters Use of openfoam for computation 7/29 3 min on a classical computer Source : Sirehna

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Population evolution in space objectives F1 : Minimisation of the pressure loss between inlet and outlet F2 : Minimisation of the standard deviation of the output speed profile Random initialization, NSGA-II, 20 generations, 32 individuals (640 solutions evaluated)~=2hr computation. All computations run from Scilab. Source : Sirehna 8/29

02 FROM 1D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Pareto front found and associated pressure profiles Source : Sirehna 9/29

02 FROM 1D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Pareto front found and associated speed profiles Source : Sirehna 10/29

02 FROM 1D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Simple writing of function //-------------------------------------------------------------------------- //--- EGO ------------------------------------------------------------------ //--- Run EGO using R and load new point ------------------------------- if verbose then R_executable_disp = " > " + optim_directory + "R_log_" + sprintf("%03d",1) + ".txt"; R_executable_disp = R_executable_disp + " 2> " + optim_directory + "R_log_" + sprintf("%03d",1) + "_Error.txt"; else R_executable_disp = ""; end unix(r_executable+" < run_ego_quantile.2.r --no-save --args " + optim_directory + " " + R_executable_disp); for j = 1:nOptimStep disp("--------------------------------------------------------------"); disp("----------------------------- "+sprintf("%03d",j)+" ----------------------------"); xnew = fscanfmat(optim_directory + "xnew.txt"); eqi = fscanfmat(optim_directory + "eqi.txt"); qk = fscanfmat(optim_directory + "qk.txt"); qmin = min(qk); //--- Compute distance between new point and existing points ----------- dist_to_px = zeros(1, nobs); for i = 1:nobs dist_to_px(i) = norm(px(i,:) - xnew); end [mindist Iclosest] = min(dist_to_px); //--- Run new experiment ----------------------------------------------- if mindist > 1e-2 //--- Case 1: xnew does not exist already -------------------------- iobs = nobs + 1; // Run simulation at new point directory = optim_directory + "xp_" + sprintf("%03d",iobs) + filesep(); Xnew = xnew.*(ub - LB) + Xbar; select problemname case "Test" ynew = test7d(xnew) + normrnd(0,alltau2(1)); case "OMD2Case1 Source : Sirehna 11/29

02 FROM 1D CASE TO FULL 3D INDUSTRIAL OPTIMISATION 3D case HVAC pipe Objective of test case: Test of the distant computation Test of the partial convergence method Objective of the problem : Similar to the previous one : find an optimal geometry (8 parameters) 25 min for the evaluation of one solution (300 000 cells) => Large computation times impose to use HPC for optimisation Source : Sirehna 12/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION 3D case HVAC pipe Software used : Geometry generated with Catia Mesh generated with StarCCM+ Solver : OpenFoam Use of ProActive as Middleware Distributed computations : Use of HPC PACAGrid (Sophia Antipolis) Use of Catia as a web service (Nantes) Source : Sirehna 13/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION OMD2 User SCILAB PLATEFORM Parameter files XML Jobs Interface Scilab / Proactive Results PACA GRID SIREHNA INRIA Server SIREHNA Server Web Server DisCalWS Dataspace PACA Grid Dataspace CATIA 14/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Example of a remote use by WUT since january 2011 (Warsaw University Technology) WUT CAD parameters Local open source middleware CATIA Webservice CAD geometry HPC PACA Grid (1200 cores) 15/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Population evolution in space objectives F1 : Minimisation of the pressure loss between inlet and outlet F2 : Minimisation of the standard deviation of the output speed profile Random intialization, NSGA-II, 20 generations, 32 individuals (640 solutions evaluated)~=11days computation on a single CPU Source : Sirehna 16/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Objective of test case: Demonstrate the use of a multi-level modelling on an industrial test case, with use of distributed computation Objective of problem From an initial ship hull without bulbous bow, find an optimised ship hull with a bow that minimize the forward resistance and decrease motions at sea Source : Sirehna 17/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION 3D case Ship hull optimisation (Cochise project) Two objectives : Reduce ship resistance for a speed of 10 kt Reduce acceleration at the rear point ( Seakeeping criteria ) Constraints : Length of the ship Length at waterline Displacement Longitudinal CoG (LCG) Metacenter (GM) Geometrical parameters : 6 global parameters + 4 parameters for the bulbous bow Cochise: Project partially funded by French Fisheries Organisation Source : Sirehna 18/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Numerical facts Mesh : 1.5e6 cells for a half-hull Use of 12 cores with PACA Grid for CFD computation (StarCCM+) One CFD computation : 10 hrs => No optimisation can be performed directly => Develop a set of strategies to obtain results with the less computation time Modelling Simulation Optimisation Computation Source : Sirehna 19/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMISATION Multilevel optimisation strategies on ship hull (COCHISE project) : Software tools used: CATIA REVA AQUA+ - STARCCM+ Use of modefrontier to perform optimisation Multilevel modelling : Use of potential code and viscous code (RANS equation) Automatic stop of simulation based on convergence criteria Restart of new solution with mapping techniques ( Closest solution fields from previous computed hulls are used to initialize the new solution field) Multi-model approach : Optimisation of hull without bulbous bow in a first time, then optimisation with the bulbous bow Use of surface response meta-model to increase optimization speed Source : Sirehna 20/29

02 FROM 2D CASE TO FULL 3D INDUSTRIAL OPTIMIZATION Objective space view Each point represents a design / solution in the objective space -25% - 4.5% START Add of a bulbous bow : Reduction of 25% on the forward resistance Source : Sirehna 21/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW OMD2/CSDL (Complex Systems Design Lab) projects collaboration Will be available on Scilab forge http://forge.scilab.org/index.php/p/omd2/ Private project up to first release Scilab Optimization Platform Batch mode (script edition, large scale execution) GUI mode (interactive edition, prototyping) Will become Scilab external module available through ATOMS 22/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW Platform features Platform organised in different modules - Data management - Modelling - Optimisation - Visualisation Wrapper - Scilab algorithms - External tools - ProActive Mask complexity for users Source : Digiteo 23/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW SCILAB Platform MMI for interaction (Definition of strategies, parameters, ) - Simulation (Number of cells, iterations, ) - Optimisation (NSGA2, ) - Modelling (DACE, Lolimot, ) MMI to build test case -Design parameters (X1,X2,X3, ) -Responses (F1,F2,F3, ) DOE X F(X) Sensitivy analysis Metamodelling Optimisation API SCILAB -> Proactive X Test case F(X) API Proactive -> SCILAB Cas test 2,3,4,5, ProActive XML Job -CATIA -CCM+ -OpenFOAM -Pretreatment / Posttreatment Execution on PACAgrid 24/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW Data management module Factors/Parameters: - Load existing Design Of Experiments (Isight.db files, ) - Generate Design Of Experiments: - DoE generator wrappers (lhs, ) - DoE generator settings Responses simulation using: - External tool (openfoam, Catia,CCM+, ) - Scilab function 2D visualization: - Factor / Factor - Response / Factor Source : Digiteo 25/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW Modelling management module Point selection: - Learning points used for modelling - Validation points used to validate model - Bad points (simulation issue, ) Modeler: - Selected among modeler wrappers (DACE, Lolimot,...) - Parameters configuration - Multiple model management with best model user selection Visualization: - 2D models - Cross correlation - Sensitivity analysis Source : Digiteo 26/29

03 SCILAB PLATFORM A REMOTE NUMERICAL WORKFLOW Optimization module Responses coefficients values setting Optimize - Selection among generic wrappers (optim, fmincon, genetic algorithms, ) - Optimizer configuration - Enable two chained optimizers Visualization - Optimal point - Pareto fronts - Robustness Source : Digiteo 27/29

04 OUTLOOK DELIVERABLES, WORK IN PROGRESS, DISSEMINATION OMD2 3D CAD parametric modelling & optimisation Automation of distributed tool chain Development of specific and appropriate algorithms and strategies dedicated to large optimisation problems Scilab Use of open-source software in industrial context Fast integration of optimization algorithms Integration of test cases in the platform All developments are to be included in an ATOMS module (End of 2012) 28/29

Thanks for your attention!! 29/29