[gmx-users] Error using double precision

[Mark Abraham]

edvin.erdtman at nat.oru.se wrote:
>>> Ok (you didn't say that in the first mail...).
>>> Does it also happen on one processor?
>> I Haven't tested in d.p.
>>
>>> After how many steps does it happen? (see Marks mail).
>> after 300000 steps, that is 600 ps.
>>
>>
>> I energy minimized the system in d.p. and it is still running after 
>> 710 ps. 
>>
> 
> No had stopped at 710 ps...
> 
> Why is it then working with single precision?

You have the possibility of observing different behaviour because the 
simulation is of a chaotic system. Small changes in starting conditions 
lead to radical differences later in time. If your starting 
configuration has some "unhappiness" that has to relax (as is normal for 
arbitrarily-solvated biomolecules) under one numerical regime it might 
explode instantly, under another it might start to relax but some 
resonant effect leads to problems later, and under yet another it finds 
the right equilibrium ensemble without breaking. The way to avoid this 
behaviour is to follow a regime like I recommended in the URL I gave 
last time. If that is not gentle enough to massage the system into good 
behaviour (when the topology is correct) then there are further degrees 
of gentleness you might apply.

Mark