[gmx-users] progressive imbalance in REMD

[Mark Abraham]

On 14/04/2012 1:46 AM, francesco oteri wrote:
> Dear gromacs users,
> I am running REMD through gromacs 4.5.5 using 10replicas.
> I am experiencing a problem with simulation efficiency, in particular 
> from gromacs output, like the following:
> vol 0.49  imb F  4% vol 0.64  imb F  8% vol 0.17  imb F  2% vol 0.56 
>  imb F 10% vol 0.17! imb F  7% vol 0.75  imb F 11% vol 0.45  imb F 11% 
> vol 0.13! imb F  8% vol 0.45  imb F 16% vol 0.55  imb F 23% step 
> 735900, will finish Mon Apr 16 07:29:53 2012
>
> it seems that higher temperature replicas suffer of an higher 
> imbalance between force and PME.
>
>
> These are the average values:
>
> 4.58991117815
> 5.5175129881
> 6.32679738562
> 7.21887045416
> 8.1979219038
> 9.45466733702
> 10.9115133233
> 12.5899111781
> 15.0987095693
> 19.5630970337
>
> Of course this problem impacts on overall performances.
>
> My questions are:
> 1) Is the progressive imbalance expected?
> 2) Is there any way to alleviate the problem?

Guessing wildly in the absence of a description, you're running NPT 
REMD, and so the particle density changes with T, so the nonbonded cost 
varies with T while the PME cost does not. The timing breakdown at the 
end of the individual .log files may prove informative in this respect. 
This problem snowballs - your generalized ensemble can only progress at 
the rate of your slowest contributing ensemble. In theory, one could 
develop a scheme where the PME performance and accuracy was 
near-constant with respect to T by varying the cutoff, splitting 
parameter and Fourier grid, but since most people choose their PME 
parameters by copying people who pulled near-arbitrary numbers out of 
the air, this would seem to be overkill.

People often do NVT REMD to avoid this effect if they are interested 
only in the ensemble at one temperature. That means the 
higher-temperature replicas have unphysically high pressures, which 
might or might not prove to be useful for enhanced sampling. Some people 
think that makes the sampling at the low temperature bogus, but I have 
never seen a convincing argument that all the replicas should correspond 
to a physical ensemble that closely resembles the target ensemble.

Mark