Improving performance
The overall performance depends on many factors, and it is therefore not possible to come up with a set of options that will always maximise performance. However, here we provide some guidance and list some tips to help you improving the efficiency of your simulations.
First of all, note that at the end of each simulation (if the code was not compiled with the -DMOSIX=On CMake option) oxDNA will print information about the total running time, the time taken per simulation step (in milliseconds) and a detailed list of timings. Here is an example:
INFO: Total Running Time: 10.6746 s, per step: 1.68876 ms
INFO: Timings, in seconds, by Timer (total, own, spent in children)
> SimBackend 10.675 (100.0%) 0.023 ( 0.2%) 10.651 ( 99.8%)
***> Lists 3.170 ( 29.7%) 3.170 ( 29.7%) 0.000 ( 0.0%)
***> Thermostat 0.005 ( 0.1%) 0.005 ( 0.1%) 0.000 ( 0.0%)
***> Forces 7.055 ( 66.1%) 7.055 ( 66.1%) 0.000 ( 0.0%)
***> First Step 0.401 ( 3.8%) 0.401 ( 3.8%) 0.000 ( 0.0%)
***> Observables 0.020 ( 0.2%) 0.020 ( 0.2%) 0.000 ( 0.0%)
Each row below SimBackend presents the time spent by doing a specific task in seconds and, between brackets, in percentage. The list of possible tasks (which depend on the specifics of the simulation) is:
Name |
Type |
Task |
|---|---|---|
Lists |
All |
Building of the neighbour list |
Thermostat |
MD |
Thermostatting the degrees of freedom |
First step |
MD |
First step of the velocity-Verlet integration |
Forces |
MD |
Computing pair interactions + second step of the velocity-Verlet integration |
Hilbert sorting |
MD on CUDA |
Sort the particles to optimise data access on GPUs.Used only if |
Rotations+Translations |
MC |
Monte Carlo rotations and translations |
Volume Moves |
MC |
Monte Carlo Volume moves |
Observables |
All |
Computing and printing simulation outputs (energy, configurations, distances, etc) |
Given a certain task, the time spent by the code is split into total time, time spent by the task itself, and time spent by the task’s children (subtasks, if you will). Note that the latter two sum up to the former.
The most important information, optimization-wise, is provided by the percentage column.
Molecular dynamics
In molecular dynamics simulations, the two most important parameters, performance-wise, are verlet_skin and dt.
verlet_skincontrols the distance a particle has to move to trigger the update of the neighbour lists. Using a larger value results in fewer updates but a larger number of possibly-interacting pairs of non-bonded particles and vice versa. You can try to decrease this number if the simulation tends to spend most of its time computing pair interactions. The most common optimal value lies between 0.05 and 0.2.dtis the integration time step. Given a number of simulated time steps, the simulated time is proporional todt. Since the overall performance of the simulation is only weakly dependent on the integration time step, using the largest possibledtvalue is advisable. However, using a value that is too large will result in numerical instabilities. Depending on the thermostat settings, optimal values range from 0.001 to 0.003 but your mileage may vary, so feel free to experiment withdtif you want to really optimise your simulation.
GPU simulations
When running CUDA-powered simulations, the box size has a non-trivial effect on performance, and its interaction with other parameters such as salt_concentration, verlet_skin, and possibly others, make it hard to come up with a way of automatically set the best options for a given case.
Since there is no dynamic memory on GPUs, in order to avoid crashing simulations oxDNA sets the size of the cells used to build neighbouring lists so that their memory footprint is not too high. If you want to optimise performance is sometimes worth to set cells_auto_optimisation = false so that oxDNA uses the smallest possible cells (at the cost of memory consumption). If the resulting memory footprint can be handled by your GPU you’ll probably see some (possibly large) performance gains.
There are some heuristics that attempt to limit the memory consumption of CUDA simulations. First of all, the given combination of parameters is used to evaluate the minimum size of the cells required to build neighbouring lists, \(r_m\). In turn, \(r_m\) is used to compute the number of cells along each coordinate \(i\) (where \(i = x, y, z\)) as
where \(L_i\) is the length of the box edge along the \(i\)-th direction. This value of \(N_i\) is the number of cells used for the simulation if cells_auto_optimisation is set to false. However, if it set to true, which is the default, then the code checks whether \(N_i > \lceil f L_i \rceil\), and if it is then sets
where
The maximum number of particles that are in each given cell, \(M\), is another important parameter that can be, to some extent, tuned to avoid crashes. It is defined at the beginning of the simulation, and also each time the total number of cells changes while the simulation is running, as
where \(f_\rho\) is a factor that can be set with the max_density_multiplier option and defaults to 3, while \(M_\text{max}\) is the number of particles found in the cell containing the largest amount of particles in the current configuration.
Note
On newer versions of oxDNA (> 3.6.1), setting debug = true will report in the log file (or on screen if log_file is not set) the amount of memory that is requested by each allocation on the GPU.
Monte Carlo
When running Monte Carlo simulations the efficiency of the sampling depends on specific moves employed during the simulation. For regular Monte Carlo and VMMC simulations, the most important options are delta_translation and delta_rotation, which set the maximum displacement for translations and rotations. Optimal values depend very much on the system at hand, so it is hard to provide some guidelines, although often values around 0.1 given decent performance. Sometimes it may be worth to set adjust_moves = true (together with equilibration_steps > 0) to let the code look for optimal values.