CUDA Vicsek model simulator

For details on the model and usage of the different types of noise refer to: Clusella and Pastor-Satorras, Chaos 31, 043116 (2021). Cite this work if you use this or parts of this code. This software is for the Vicsek model defined on Euclidian space only.

External tools

The software makes a fundamental use of thrust in two cases:

To compute the average polarization of the system (function compute_phi in file system.cu)
To sort the box list devBox with thrust::sort_by_key in vicsek.cu.

Also:

We use GNU Scientific Library (GSL) to compute moving window statistics of the order paremter phi. This is not fundamental and can be easily changed in the vicsek.cu file.
Most output files can be easily plotted using gnuplot, but any other plotting software can work provided it can read binary files.

Compilation

From the source folder, the code can be compiled with

nvcc system.cu vicsek.cu -O3 -lcuda -lcurand -lgsl -lgslcblas -o ../gpuvs

System parameters and input

Upon compilation the program can be executed from the workspace with

./gpuvs <flag> <L> <rho> <eta>

where

flag is a string to identify output files
L is the lattice size
rho is the density of particles
eta is the noise strength. Notice that different types of noise are impletented in the code, the user needs to comment/uncomment them manually in the update_bitxus() function at system.cu.

Particle number N is directly computed from rho and L. The box size for easy particle location is fixed to be 1.

A typical call to the program could be

./gpuvs_v1 experiment_A 128 2.0 0.62

Other system parameters need to be changed directly in the code:

The simulation time Tf and transient trans are defined in the main() function in vicsek.cu. This is, by default, Tf=2.5e5, trans=5e4.
The distance radius for the nearest neighbour computation is the macro R0 defined in system.cuh. However this should not be increased to more than 1 without changing some important parts of the code. Otherwise one might need to look at more than 8 neighbouring boxes to compute the local polarization, and this is hardcoded.
The constant particle speed is the macro V0 defined in system.cuh.

Workspace structure and output files

Any of the following can be easily adapted modifying the source code vicsek.cu.

Output files are stored in three different folders, that need to be created before:

timeseries stores the binary file containing the time series of the order parameter
outputs stores the final statistics of phi in plain format
snapshots stores the position and velocity of each particle in binary format

The output files are authomatically named with the provided values of L, rho, and eta, and also with the "flag" string passed to the program.

Using gnuplot one can plot the time series of phi with:

plot 'timeseries/<filename>.bin' '%double' using 0:1 with lines

The snapshots of the system can be also plotted using, for instance,

plot 'snapshots/<filename>.bin' binary format='%double%double%double%double' using 1:2:(atan2($4,$3)) with dots lc palette

Algorithm

The program is based on the usage of bitxus in boxes, i.e., cell lists.

A Bitxu is a struct defined in system.cuh that contains all the information of each particle of the Vicsek model, namely, its box location, position, velocity, and local polarization. The word bitxu comes from a direct (and wrong) spelling of the spanish word "bicho" (bug) in catalan pronunciation. If you are confused by the word, you can think that a bitxu is a particle, bird, fish, bacteria, sheep, or whatever your are trying to model.

A box is not explicitly defined in the code, and is one of the L^2 squares of unit size in the lattice. This is the cell in cell lists.

The Vicsek model is an array of "bitxus", that in our code is called devNodes. Each bitxu is at each time in one and only one box. Knowing what box a bitxu belongs to is easy to compute, and it is at all times stored in devNodes[i].idbox.

However, since bitxus move in time, knowing how many and which bitxus are in a given box at a specific moment is difficult. This is the main problem to be solved. To address it I create a list called devBox. Counterintuitively, this is not a list of boxes, but a list of bitxus'. This list needs to be sorted at all times. Therefore, in a system with 10 bitxus and 4 boxes, devBox could look like

devBox=[ 0 0 1 1 1 2 3 3 3 3]

where it is clear that we have 2 particles in box 0, and 3 in box 1, 1 in box 2 and 4 in box 3.

Since this list is sorted at all times, in order to access the elements of devBox easily we can just define two arrays, called devIni and devEnd, of size L^2, i.e., the total number of boxes. These vectors respectively indicate the initial and final index of the particles in a given box of devBox.

Therefore, in our previous example,

devIni=[0 2 5 6]

and

devEnd=[1 4 5 9]

since the first bitxu of box 0 is referenced at devBox[0] and its final bitxu is referenced at devBox[2], whereas for box 2 we only have one subject which is at position 5, and therefore devIni[2]=devEnd[2]=5.

When the array devBox changes, devIni and devEnd need to be updated. This is done by the function get_positions(), defined in system.cu. Thanks to this scheme, it is easy to retrieve all the particles in a specific box. To compute the local polarization of a particle one just needs to look at the particles in its own box and the surrounding boxes. This is done by the function compute_distances().

At any time steps, bitxus change boxes, and therefore, the array devBox needs to be updated. This is done in two steps.

When particles move, they update their position in devBox. This is done in the function update_bitxus() in system.cu (where devBox is called boxlist). However, now devBox is not sorted.
We call sort_by_key from thrust to sort devBox. However, now, particles in devBox are not properly identified by their index, we need to re-organize particles in devNodes the same way they have been reorganized in devBox. This reorganization is given by the array devSL. We then just need to make devNodes[i]=devNodes[devSL[i]]. This is done with the help of an auxiliary array devTemp (despite its name, this is not a temporal array at all), and the copy and swap functions.

pclus / vicsek-cuda