parallel computation

Hi Ning,

as I remember, Dion said, that it should be at least 5000 spheres on each
process to get more-less normal efficiency. You have more. But what PC do
you use for simulation? I tried to increase the number of processes on
2-core machine with no success. 4-core machine works much more better with 3
working processes + 1 master.

______________________________
Anton Gladkyy

2010/4/26 ceguo <email address hidden>

> New question #108447 on ESyS-Particle:
> https://answers.launchpad.net/esys-particle/+question/108447
>
> Hi everyone,
>
> Initially I used a small amount particles in the simulation, so I never
> tried parallel computation before in ESyS, (i.e. I set
> numWorkerProcesses=1,mpiDimList=[1,1,1] in the script), though it is a great
> advantage characteristic of ESyS. Now I want to increase the number to
> 40,000 in an isotropic compression test and I hope to use parallelism to
> improve the efficiency of the simulation. But as I tried some effort, when I
> increase the numWorkerProcesses(trying 2, 4 and 8 respectively), the
> computation time increases obviously at the beginning. I haven't continued
> to run the whole simulation as I found the first few steps slowed down. I
> know the communication between domains will cost time. But I think 40,000 is
> not a small amount. So I am wondering how many particles do you usually have
> when you consider to use parallelism?
>
> Rgds
>
> Ning
>
> --
> You received this question notification because you are an answer
> contact for ESyS-Particle.
>

Hi Ning,

I routinely run simulations of approximately 50,000 particles and typically use 4 worker processes for those. For uniaxial compression simulations, the total runtime is about 12-20 hours (for about 250,000 timesteps).

Anton is quite correct that you will need to ensure each worker process is responsible for approximately 5000 particles or so. This will ensure a reasonable trade-off between computation time and communication time.

I've also found that using more worker processes (+ master) than you have processors/cores will result in very bad performance. Since each worker process uses almost 100% of the CPU, sharing workers on the one CPU will at least halve your performance.

Cheers,

Dion.