MolSim

The Molecular Dynamics teaching code.

Group H

Members:

• Samhitha Girish Jois
• Katharina Miller
• Anna Lena Müller

Code

• Link: https://github.com/Kotronon/Bachelor-Praktikum
• Branch: master
• Revision: 4ba3a23
• Compiler: gcc 11.4.0

Run Code

• to run the program:
  • ./MolSim
  • due to the xml input being incomplete it is only possible at the moment to change parameters in the main method manually as well as create cuboids/disks there
  • this should be fixed in the future, log level can not be chosen either currently as this was planned to be set up completely in the xml files
• to run the tests:
  • ctest

Report

Task 1

• Simulation of a Membrane
  • Worked on by Samhitha, report separate

Task 2

• Parallelization
  • Only one strategy mentioned, other one was worked on by Samhitha
• Performance analysis
  • After analyzing the performance of the Rayleigh-Taylor simulation in 3d with Perf it is very noticeable that the vast majority of time is spent in the force calculation
  • The LennardJonesForceCell function takes up nearly 94% of the calculation time
  • Other calculations with a simple iteration over all particles only take up < 1% of the time
  • -> Due to thread creation taking up time as well the first strategy only focuses on optimizing the force calculation
• Parallelization strategy
  • Parallelize first loop in LennardJonesForceCell, each thread gets assigned a group of cells
  • Each thread gets neighbours of this specific cell and calculates forces between particles in the same cell and neighbour cells
  • Due to Newtons third law being used during the calculations the particle updates have to be considered critical zones to avoid race conditions
  • Due to other calculations/functions being simple loops the thread creation usually takes up more time than it speeds the simulation up
  • -> Because of this only the force calculation is parallelized
• Possible improvements:
  • Depending on the kind of simulation it might be beneficial to add information on what scheduler to use
  • If number of particles is uneven across the cells threads might not be balanced, but schedulers to solve this also take more time
• Speedup of Parallelization:
  • Speedup was calculated as average over 1000 iterations of the 3D Rayleigh-Taylor simulation
  • More iterations were too slow to run on 1 thread in a reasonable time
  • Specs:
  • System: Kernel: 5.15.0-88-generic x86_64 Desktop: Gnome 3.38.4 Distro: Zorin OS 16.3 base: Ubuntu 20.04 LTS Focal
  • CPU: Topology: 8-Core model: AMD Ryzen 7 5825U with Radeon Graphics
  • Conditions: measuring of time duration with functions in chrono library, disabled I/O operations
  • Results:
  • Result as SpeedupParallelization.png in submission
  • Speedup is steadily increasing with number of threads, reaching its peak on 8 threads with a speedup of about 3.64
  • After this speedup is decreasing again due to the pc this was run on only possessing 8 cores
  • Threads can not be run truly in parallel and the overhead of thread creation and management slows down the simulation more

Task 3

• Rayleigh-Taylor instability in 3d
  • Simulation could not be run completely due to simulation time not being reasonable (>70h)
  • This is most likely due to the optimization task from the last worksheet not being completed
  • Video provided in submission shows the first 20 000 iterations instead of the full 200 000
• Branch to run simulation: simulation_rayleigh_taylor_instability_3d

Task 4

• Nano-scale flow simulation (Option A - not chosen)
  • We started with this task but decided to switch to task 5
  • The walls are working and the new temperature calculations were done as well
  • But due to some issues we did not finish all the tasks
  • The uncompleted code can be found in branch Task4 (https://github.com/Kotronon/Bachelor-Praktikum/tree/Task4)

Task 5

• Crystallization of Argon (Option B - chosen)
  • The new force calculations had to be added
  • While computing the simulations we noticed multiple mistakes in the 3D computations
  • To solve this we needed to add more ghost particles and make sure all functions in LinkedCellContainer work properly in a 3d space
  • We improved our LinkedCell structure like you suggested by changing the cell sizes accordingly
  • Equilibration: gets gaseous after 8 seconds in our video. That equals 30000 iterations
  • Cooling: begins to get solid after 25 seconds in our video. That equals 125000 iterations
  • Supercooling: gets solid after 6 seconds in our video. That equals 25000 iterations
• Branches to run simulations:
  • simulation_argon_equilibration
  • simulation_argon_cooling
  • simulation_argon_supercooling

Misc

• We needed to comment out the thermostat tests because it suddenly would not work anymore, and we could not find the mistake in time
• All other test worked perfectly fine, so we were quite confused since the temperature in our simulations seemed to be calculated correctly