Profiling of Modelica Real-Time Models Christian Schulze 1 Michaela - PowerPoint PPT Presentation

GI Profiling of Modelica Real-Time Models Christian Schulze 1 Michaela Huhn 1 Martin Schüler 2 1 Technische Universität Clausthal, Institut für Informatik, Julius-Albert-Str. 4, 38678 Clausthal-Zellerfeld, Deutschland {Christian.Schulze | Michaela.Huhn}@tu-clausthal.de 2 TLK-Thermo GmbH, Hans-Sommer-Str. 5, 38106 Braunschweig, Deutschland M.Schueler@tlk-thermo.de

GI Overview 1. Introduction 2. Profiling of Modelica Models 3. Implementation 4. Case Study 5. Conclusion Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 2

GI Introduction  Usages of RT simulation: - Rapid Control Prototyping (RCP) - Hardware-in-the-Loop (HiL) - Model Predictive Control (MPC)  Hard real-time (execution time per global solver step <= output step size)  Overruns due to events and algebraic loops  Model has to be improved manually  Profiling helps to trace back the cause of an overrun  There is no applicable profiling tool  We want to profile on the RT-Target: - Calls to external functions (libraries) - Algebraic loops  We want minimize the overhead caused by profiling Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 3

GI Overview 1. Introduction 2. Profiling of Modelica Models 3. Implementation 4. Case Study 5. Conclusion Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 4

GI Simulation on Real-Time Targets  To execute a model on a Real-Time target one has to: Convert to C-source (export) Compile the model application Transfer to the real-time target Execution Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 5

GI Profiling  Tracing vs. Profiling - Call-Graph - Flat-Profile  Instrumentation must be added but causes additional work load  We are profiling by measurement (just like Tau) but are not logging the callee or point in time  One can draw this information from the general structure of those models Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 6

GI Source code of Modelica models  Separated into Initialization and Simulation  Solver step: - n I integration steps - e I external function calls - c I additional calculations - a I (non-)linear blocks - e aI external function calls - c aI additional calculations - 1 output of variables - e O external function calls - a O additional calculations  flat profiling for each section  profiling is done separately for each solver step Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 7

GI Overview 1. Introduction 2. Profiling of Modelica Models 3. Implementation 4. Case Study 5. Conclusion Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 8

GI Export of models Model created in Simulator Export Solver Model converted to C source files Target I/O-methods User source code C source files of the model application Compiler Application Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 9

GI Modified Export of models Model created in Simulator Export Solver Modified model source files Target I/O-methods User source code Profiling I/O-methods C source files of the model application Compiler Application Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 10

GI Modified Source Code  Major impact on the overhead: - Saving profiling data (access to HDD)  Access to HDD is outsourced to secondary application  Model application is a Real-Time Kernel Task(SCALE-RT) - Prioritized execution  Model runtime increases at maximum by 4% Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 11

GI Implementation  Communication between user task and model task Real-Time Linux Kernel Space (RTAI-API) User Space Model User Space (Kernel module) Buffer FIFO Application Real-Time Task data exchange data storage Source code HDD Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 12

GI Overview 1. Introduction 2. Profiling of Modelica Models 3. Implementation 4. Case Study 5. Conclusion Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 13

GI Case Study  Steady State Continuity Equation Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 14

GI Model runtime over simulated time 647 µs 547 µs model runtime 447 µs 347 µs 247 µs 147 µs 47 µs -0,2s 0,0s 0,2s 0,4s 0,6s 0,8s simulated time Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 15

GI Results of Profiling  Global solver step separated into integration steps and output of results integration integration steps; 47,61µs steps; 47,61µs integration integration steps; 47,61µs steps; 47,61µs output of results; 5,34µs overhead by global solver; 7,76µs Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 16

GI Results of Profiling  Integration steps separated into external fluid property calculations, non- linear equations and overhead integration integration steps; 47,61µs steps; 47,61µs integration integration steps; 47,61µs overhead; steps; 47,61µs 3,240 fluid non-linear properties; output of equations; 8,052 results; 5,34µs 36,290 overhead by global solver; 7,76µs Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 17

GI Algebraic loops  Closer look on integration steps non-linear fluid equations; properties; 36,290 8,052 overhead; 3,240 Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 18

GI Algebraic loops  Non-linear equations assinged to submodels tube.Cell.heatPort.Q_flow_0; 0,53 tube.Cell.heatPort.T_1; 0,518 _der_tube1.Cell.T_1; 1,282 _der_tube.Cell.T_1 _der_tube1.Cell.T_0; 1,3 pump2ndOrder.P_shaft _der_pump2ndOrder1.T; 1,408 _der_pump2ndOrder.T pump2ndOrder1.P_shaft; 0,566 _der_tube.Cell.T_0 tube1.wallCell.T_1; 0,506 pump2ndOrder.v_flow tube1.wallCell.portS.T_0; 0,508 ; 28,27 sink1.port.h; 0,722 pump2ndOrder1.v_flow; 0,68 non-linear fluid equations; properties; 36,290 8,052 overhead; 3,240 Steady State continuity equation Dynamic State continuity equation Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 19

GI Mass flow of inlet and outlet  Error due to thermal expansion of liquid dynamicSC_source dynamicSC_sink steadySC_sink&source 0,3158 mass flow rate [kg/s] 0,3156 0,3154 0,3152 0,315 0,3148 0,3146 0,3144 0,3142 0,314 0,3138 2 2,5 3 3,5 4 time [s] Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 20

GI Overview 1. Introduction 2. Profiling of Modelica Models 3. Implementation 4. Case Study 5. Conclusion Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 21

GI Conclusion  Saving results to HDD should be outsourced to a secondary non-Real-Time application  Flat profiling by measurement for each step gives more than enough information  Profiling helps identifying the work load contributions and aids the user optimizing the model Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 22

GI Thank you for your attention! Christian Schulze Institut für Informatik EOOLT 2010- Profiling of Modelica Real-time Models 23