Using Neural Network-based Approximation to Improve Performance of HPC Applications

In this project, we use neural network models to approximate computation of an HPC application to improve performance (reducing execution time) of the application. The neural network should use the same input and output variables as the HPC application.

To use a neural network model to approximate computation, we face multiple research challenges.First, we must make sure that our neural network models can bring a performance benefit. This means the execution time (inference time) of the neural network model should be shorter than that of the original HPC application. To address this issue, we will try different neural network models based on the performance (execution time) of the original code to decide which model should be used. Using such a performance-driven approach to select the model separates us from the existing work where the model accuracy is often used as the only metric to select the model.

Secondly, we must determine appropriate input and output variables for the neural network models. Those variables are also the input and output variables of the approximated application.

We will study a specific HPC application, openFuelCell in this project. The openFuelCell is a popular code for computational fluid dynamics (CFD) to model fuel cells (http://openfuelcell.sourceforge.net/project). We want to improve performance of openFuelCell without losing simulation accuracy.

In the openFuelCell project , you can find and run the sofcFoam model using the instructions in “http://openfuelcell.sourceforge.net/doc/quick-start”. Use scripts under “/openfuelcell/run/ crossFlow” to run your openFuelCell code. To use this scrip to run crossFlow, you will need to OpenFOAM version 6. See https://openfoam.org/download/6-ubuntu/ for details.

Our Contribution :

• Analyze openfuelcel :
  • Profiling : tool used --- gprof
  • Identify Memory Leak : tool used --- valgrind
  • Trace Code Stack : tool used --- gdb
• Collecting DataPoints :
  • openfuelcell Models used for datapoint generation : coFlow, crossFlow, quickTest, quickTestStack
  • Parsing data : Used a python based script for parsing data
• ML Model Implementation :
  • MLP : In Progress.

Notes :

• Based on profiling result it was difficult to identify most time comsuming part of the code. Basically all of them are operator assignment . Later we dicided to approximiate total comuputation of the model. Now, our terget variable is voltage,current.

• During datapoint collection we were facing floating point exception error after certain timesteps. For quickTestStack(2614) we were able to get highest number of datapoint among all model’s. For other model’s (coFlow, crossflow, counterFlow) single iteration of timestep takes 5 minutes – 30 minutes. For coFlow model it took more than 3 hours to collect 120 datapoints.

References :

[1] http://openfuelcell.sourceforge.net/
[2] https://github.com/Cantera/cantera
[3] http://courses.washington.edu/mengr331/