what is the optimal way to run parallel

LAMMPS LAMMPS Mailing List Mirror
1

Hello,

I performed some test run to see how many processor should I use for my problem. Following is what I observed:

number number processor
of of grid
processors dynamic
steps in
an hour

4 18800 1x2x2
8 27750 2x2x2
12 34250 2x2x3
16 39200 2x2x4
#20 30500 2x2x5
24 43350 2x3x4
#28 42050 2x2x7
32 47200 2x4x4
#36 43450 2x3x4
40 50600 2x4x5
#44 45950 2x2x11
#48
#52 39350 2x2x13
#56 40700 2x4x7
60 48500 3x4x5
64 50450 4x4x4

a simple fit reveals: time/step ~ 1/(sqrt(# of processors))

However, not all choices of number of processors is a good choice, e.g. using 20 processors takes longer than 16 processors.
Wondering if there is way I can choose processor grid and if this is related to time/step ? Is there also some relationship between total
number of atoms and max number of processor that could be used ? The above calculations were performed on 464 atoms.
I have also attached time/step (in hours) vs number of processors plot with this email.

Please let me know.

Thank you,

Vivek

power6.jpg

2

Vivek, benchmarking with a fixed number of atoms is one good way show your parallel performance. However, your atoms per processor is very small. For good scaling, each core should probably have several hundred atoms, or more. I usually choose a processor grid that is roughly cubic, and that is divisible by the number of cores/node.

Tim

3

Vivek:

You can specify a grid using the “processors” command. However, it looks like you’re wasting a lot of time—a simulation with just 464 atoms should run much faster than 50K steps/h! (Did you take a look at the breakdowns on where the time is spent? I’d suspect a very substantial portion of the time is wasted on communication!)

—AEI