Optimizing configuration option for fast calculations

From: DimitryASuplatov (genesup_at_gmail.com)
Date: Tue Apr 28 2009 - 07:16:00 CDT

Hello,

I have been recently writing about namd being very slow for
calculations of my 110000 atoms system.

I have received lots of suggestions and links in reply. So I was able
to adjust my configuration options that made calculations a lot
faster.

I am writing this in case someone like me has the same speed problems
and feels himself on the edge.

1/ # Force-Field Parameters
exclude scaled1-4
1-4scaling 1.0
cutoff 10.0 #(Valid with the condition of using PME)
switching on
switchdist 8.0 #Cutoff-2
pairlistdist 12.0 #Cutoff+2
stepspercycle 10

2/ # Integrator Parameters
timestep 2.0 ;# 2fs/step
rigidBonds all ;# needed for 2fs steps
nonbondedFreq 1
fullElectFrequency 2

3/ PME yes
PMEGridSizeX 128 # 128 = 2^7
PMEGridSizeY 128
PMEGridSizeZ 128

Now I have the following scaling:

CPUs seconds/step days per ns
=====================================
4 0.4 2.31
16 0.117 0.68
32 0.062 0.36
64 0.038 0.22
128 0.031 0.18
256 0.018 0.105

That makes it 24 hours for 10 ns on 256 CPUs. (At least 40% faster than before!)

Thank you all for your time and help.
SDA

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:52:40 CST