[lammps-users] about MPI and the efficiency of computing

LAMMPS LAMMPS Mailing List Mirror
1

Dear,

I have a Centos 5.1 OS computer which have 2CPUs and 4cores per CPU. I have installed MPI and started rsh,rexec service.

When I tested the MPI, I have found that the mpirun -np >1 is lesser efficient than mpirun -np 1. What does it matter?

ONE CPU

Pair time () = 3.92188 (79.1174) Neigh time () = 0.125 (2.52167)
Comm time () = 0.148438 (2.99448) Outpt time () = 0.0820312 (1.65485)
Other time (%) = 0.679688 (13.7116)

total time ~4s

four CPUs

Pair time () = 0.976562 (4.99825) Neigh time () = 0.0253906 (0.129955)
Comm time () = 14.3477 (73.4343) Outpt time () = 0.12207 (0.624781)
Other time (%) = 4.06641 (20.8127)

total time ~20s

can some kind people give me one explaination? Thanks

2

Dear Wende,
Look at the communication time data… That is making the difference. If you look at pair time, it is 1/4 in case of 4 CPUs. It seems like your system is pretty small. Parallel efficiency is in general better for simulating larger systems, where communication time is small as comparison to calculation time.

I am not sure about the 'other time"

Regards,
Vikas

3

Dear,

wende,

I have a Centos 5.1 OS computer which have 2CPUs and 4cores per CPU. I have
installed MPI and started rsh,rexec service.

enabling rsh/rexec is a bad idea unless you have a machine that is
not connected to any network.

When I tested the MPI, I have found that the mpirun -np >1 is lesser
efficient than mpirun -np 1. What does it matter?

ONE CPU

Pair time (\) = 3\.92188 \(79\.1174\) Neigh time \() = 0.125 (2.52167)
Comm time (\) = 0\.148438 \(2\.99448\) Outpt time \() = 0.0820312 (1.65485)
Other time (%) = 0.679688 (13.7116)

total time ~4s
four CPUs

Pair time (\) = 0\.976562 \(4\.99825\) Neigh time \() = 0.0253906 (0.129955)
Comm time (\) = 14\.3477 \(73\.4343\) Outpt time \() = 0.12207 (0.624781)
Other time (%) = 4.06641 (20.8127)

total time ~20s

can some kind people give me one explaination? Thanks

a benchmark that takes a few seconds is useless.
please consider that in the case of a parallel run
you first have to setup and initialize the parallel
environment.

cheers,
   axel.