[AMBER] Implicit solvent simulation is unexpectedly slow

From: 浅井 賢 <suguruasai.gmail.com>
Date: Sun, 02 Aug 2015 17:21:43 +0900

Dear Amber users,

Several years ago, I performed a simulation with explicit solvent model,
the system contained 106,658 atoms, in pmemd.cuda.MPI of Amber 11 with 2
GPUs (No Pear to Pear).
At that time, the simulation speed was 8.52 ns/day and I felt the speed
was fair enough.

Recently, I performed a similar simulation with implicit solvent model,
the system contains only 13,549 atoms (exclude solvents from previous),
in pmemd.cuda.MPI of Amber 14 with 2 GPUs (Pear to Pear enabled).
While 1) the number of atoms in the system is significantly reduced, 2)
the simulation was performed under Pear to Pear, 3) Amber 14 improved
the performance of pmemd.cuda.MPI, I speculated that huge performance
enhancement. However, when I checked mdout file, the final performance
was 8.37 ns/day thus actually it get slower.

I'm suspecting some of my configure in `mdin` but I have no idea which.
So would someone help me to figure out?
I attached `md1.in` and `difference.txt` which contains partial parts of
mdout of explicit/implicit simulations.

Best,
Asai


P.S.
I already google "Amber GB slow" or something like that and found/checked

http://structbio.vanderbilt.edu/archives/amber-archive/2007/3195.php

but I don't really think that thread can help my situlation. :-(


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber


Received on Sun Aug 02 2015 - 01:30:02 PDT
Custom Search