AMBER Archive (2003)

Subject: Diferences in gibbs

From: Miguel de Federico (miguel_at_qorws1.uab.es)
Date: Tue Mar 18 2003 - 03:37:21 CST


 I had recently compiled amber 7 in a dual Xeon server with RedHat 8.1.
I did so with g77, an after I read that ifc generates faster
executables. I have run a test job, a free energy perturbation looking
for differences between both executables generated.

 The job that run with the g77 compiled version needed 42 hours an the
result was
 Forward: -1.07 Reverse: 1.68

 The job that run with the ifc7.0 compiled version needed 35 hours, but
the result was
 Forward: 0.25 Reverse: 0.50

 The ifc 7.0 compiled version is about 20% faster, as I had read, but
the results seem quite different from those obtained with the version
compiled with g77. I know that molecular dynamics have a random
trajectory and therefore different runs give different results (altough
if it's converged then mean values should be the same). As a the result
of a FEP calculation is trajectory-dependent if the phase space is not
enough explored, I could imagine that there would be differences between
both gibbs runs. However, is so much difference considered normal?
Should I consider a short FEP calculation is not representative and try
another test job?

  When using the test set provided with amber, gibbs has passed all
tests ecept for one:

==============================================================
cd val_to_ala; ./Run.gibbs_box
Free Energy Perturbation
diffing out.fep.save with out.fep
possible FAILURE: check out.fep.dif
==============================================================

 The out.fep.dif is as follows:

1077c1077
< Etotal = -3268.07814 Kinetic = 790.23786 Potential =
-4058.31599

---
>  Etotal  =  -3268.07814  Kinetic =    790.23786  Potential  =
-4058.31600

So it seems not serious problem, I think.

Then, have anybody any opinion about this? Should I keep using ifc? Should I run other kind of test jobs?

Thank you in advance,

Miguel.