From: Francesco Pietra (chiendarret_at_gmail.com)
Date: Mon May 28 2018 - 10:05:01 CDT
On passing, I have already alluded to the deterioration of the performance
when executing a FEP. Now I have looked at the trend in some detail for the
Unbound FEP of a ligand-protein project and my question is simply whether
this is what is expected for FEP simulations or there must be a problem
with my cluster/filesystem (NextScale, single node 36 core Intel® Xeon®
Broadwell, memory 115Gb, as the limit of scalability for the small system).
To avoid stressing the cluster, according to previous advice, I have set
5000 for all outputs, including alchoutFreq.
FF CHARMM36, organic ligand in TIP3 water (natural product, accurately
parameterized at HF/6-31G* level with also torsional fitting)
I am dealing with the steep region of the Hamiltonian, along either ten or
frwd 0.0 0.2 0.0.2
frwd 0.0 0.2 0.0.1
back 0.2 0.0 -0.02
back 0.2 0.0 -0.01
On either direction of FEP, the equilibrated performance is initially
0.0026s/step. Then, after one or more windows (apparently without
regularity) it decreases to 0.005s/step. Today I observed a peak of
decreased performance, 0.01s/step, at nearly completion of window 13 of
frwd, while later the performance raised to 0.005s/step. In no case was the
performance restored to the initial value. "back" is always slower than
"frwd", whatever happens.
Because of that, it proved hard estimating the walltime and I was obliged
to allow a very large excess.
Thanks for advice
This archive was generated by hypermail 2.1.6 : Mon Sep 16 2019 - 23:19:42 CDT