PROFASI  Version 1.5
Parallel tempering with PROFASI

Parallel tempering or replica exchange method is a very popular Monte Carlo algorithm. A set of N replicas are simulated simultaneously at N different temperatures and configurations are exchanged between different replicas with a Metropolis like probability.

Parallel tempering is implemented in PROFASI using MPI. The executable is called ParTempRun.mex, and there is no serial counterpart to this program. Typically, you would set up a population (chains, box size etc.) using a settings file or the command line. The procedure to do these things is exactly the same as in canonical MC simulations (BasicMCRun program reference), simulated annealing (siman_progref) and simulated tempering (SimTempRun program reference).

Options

Apart from all set up options (which make sense) of described in connection with canonical MC run (BasicMCRun program reference), parallel tempering accepts

What do I need for a parallel tempering run ?

Example set up

Create an empty directory and a file called "settings.cnf" in it. Put the following lines in it.

log_level 10
add_chain_pdb 1 1GB1.pdb::A,41,56
box_length 100
num_cycles 5000000
cycle_length -1
tfile T.dat
conf_write_freq 1000
rt_write_freq 1000
avg_write_freq 10000
num_therm_cycles 30000
new_obs Rg rg
new_obs ProteinRMSD hvrmsd using +HV ; struc1 1GB1.pdb:1:A,41,56 ; struc2 $::A
new_obs NativenessQ Q structure 1GB1.pdb::A,41,56

This sets up a simulation of the C-terminal hairpin of 1GB1.pdb. You will need the pdb file 1GB1.pdb for the sequence and RMSD measurements. Get it from the protein data bank (http://www.rcsb.org). Now create a temperature file called "T.dat" with the following lines.

#temperature Kelvin
370
355
340
325
310
295
280
270

The first line tells the program that temperatures are given in Kelvin units. You can also write "#temperature inverted". If this line is omitted, PROFASI internal temperature units will be assumed.

Run the program using your system's implementation of mpirun or mpiexec. Something like:

mpirun -np 8 $profasi/app/bin/ParTempRun.mex

If you have no idea what kind of temperatures to use, a good start is to use a geometric series. To do this, remove the line "tfile T.dat" in the settings file. In its place, add the following 3 lines:

tmin 274 Kelvin
tmax 374 Kelvin
ntmp 8

If you now run the program as above, it will automatically create a simple temperature set (in a geometric series) and use it. The same can be done by deleting the tfile line in the settings file and instead of adding the 3 lines above, passing them in the command line:

mpirun -np 8 $profasi/app/bin/ParTempRun.mex -tmin 274 -tmax 374 -ntmp 8

If you try to run the above example set up with 5 computing cores, it will be an error. Each replica needs to be a separate MPI process. If you run it on 16 (or any integral multiple of 8) cores on the other hand, it will not be an error. There will then be 2 (or more) replicas at each temperature, the section Multiplexing discusses this further.

Multiplexing

If you set up a parallel tempering run with 8 temperatures, and run it on 32 MPI processes, there will be 4 replicas at each temperature. At the time of temperature exchanges, a replica at a certain temperature t0 attempts to exchange temperature with another at a neighbouring temperature t1. In the current example, there are 4 such replicas at t0 and 4 at t1. The exchange attempts are done between random pairings between these two groups. Without the random pairings, the run with 32 MPI processes is equivalent to 4 independent runs with 8 MPI processes each. Because of the random couplings, any replica might exchange temperatures with any other in the entire run. This idea is called "multiplexing".

The benefits of multiplexing are, at present, unclear. So, it is turned off by default, which means that if you run 32 MPI processes with 8 temperatures, you end up with data that is equivalent to 4 independent runs with 8 processes each. But it can be turned on easily in the settings file with the line "multiplexing on", or on the command line,

mpirun -np 8 $profasi/app/bin/ParTempRun.mex --multiplexing

temperature history of a single replica

A typical ProFASi simulation is about 10 million sweeps long. Therefore it would take a lot of disk space to store the run time history at every MC sweep. Normally, the run time history is written only once every 1000 sweeps or so. But often one wants to know exactly how the temperature flowed among the different replica. That information can not be reconstructed from the configuration files or the "rt" files without re-running the simulation. For this reason, there is an option "–keep_T_history".

If the option "–keep_T_history" is given, a file "T_history" is created in the output directory of every replica. That file contains (time, temperature) pairs for every successful temperature change. So, a certain portion in that file reads:

...
9543   7
9577   8
9601   7
...

it means that the replica was at temperature with index 7 from sweep number 9543 through 9576, changed to temperature 8 and remained there from sweep 9577 through 9600 before coming back to temperature 7. This information combined from all the replicas can be used to reconstruct a complete history of temperature updates in the run.

But remember: The T_history files will be huge. Use this option, only if you really want this information!

Using the option –T_update_interval (e.g. –T_update_interval 10), one can make parallel tempering attempt temperature exchanges every 10 Monte Carlo sweeps instead of every sweep.


PROFASI: Protein Folding and Aggregation Simulator, Version 1.5
© (2005-2016) Anders Irbäck and Sandipan Mohanty
Documentation generated on Mon Jul 18 2016 using Doxygen version 1.8.2