[gmx-users] poor performance in Hemiltonian Replica Exchange

[francesco oteri]

Dear gromacs users,

I performed a Hemiltonian Replica Exchange (i.e. replica exchange where
each replica has a init_lambda=0, delta_lambda=0 and init_lambda ranging
uniformely from 0 to 1).

Since I have only ten fixed discrete lambda, I run a Temperature Replica
Exchange where, for each replica I generated a .top file with the parameter
rescaled through a
python script ( in practice I did through python the same thing gromacs is
supposed to do with the H-REM previously described). Now gromacs complained
because
every replica has the same setup, so I changed the temperatures using very
close values (300.0001K,
 300.0002K,300.0003K,300.0004K,300.0005K,300.0006K,300.0007K,300.0008K,
300.0009K)
With this setup the simulation runs fine and I expect to have similar
result.

Then I compared the results observing two phenomena:

1) In the second case exchange rate is 100%, while in the first case I have
an exchange rate close to 30%.
Does it rise  because the temperatures are too close?

2) The second setup is 3x faster!
In particular I observe an imbalance between PME and force calculation
ranging from 10% to 60%.
I tried to run each replia indipendently (a different mdrun instance for
each .tpr file) but still I observe the same performance slowdown.
I guess the free energy impairs the efficient force calculation, but I dont
understand why.

Can someone explain me the two observations?



Francesco
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20120510/f5452d69/attachment.html>