Re: Parallelisation

[Rudolf Weeber]

Hi Ahmad,
On Wed, Jan 12, 2022 at 01:35:35PM +0000, Ahmad Reza Motezakker wrote:

> 
> I have a suspension of polymers coupled with fluid. (LJ+LB)
> 
> Here are the parameters:
> 
> box_l = 300*sigma  (box is a cube)
> 
> number of polymers = 300
> 
> beads per polymer = 26
> 
> All the particles = 300*26 =7800
> 
> LJ cut = sigma*(2**(1/6))
> 
> l_skin = 8.3 *sigma (set it thid to have 31cells in each direction)
> 
> LB cells = 50
> 
> number of cells in each direction = 31
> 
> 
> Timing for 100  productive run after setting the system and warming up:
> 
> 1core     17.824 s
> 
> 2core      16.22 s
> 
> 4core      15.93 s
> 
> 8core       17.83 s
Can you please report the timings obtained via

mpirun -np 4 ./pypresso  ../maintainer/benchmarks/lb.py 
--particles_per_core=20000 --lb_sites_per_particle 6
These are 80k particles with a 78^3 LB, so slightly bigger than your system. I 
get about 80ms per time step on an AMD Ryzen 1920x Threadripper with 12 cores.
You can also check on 8 cores by using -np 8  and --particles_per_core=10000. 
On my system, this is not worth it.
You can get a significantly faster simulation by using the GPU LB. The speedup 
relies partially on the fact that GPUs are very well suited for LB (e.g. 
because of high memory band width) but also on the fact that Espresso's GPU LB 
uses single precision, whereas the CPU one uses double precision.

> If I want get one node with 128 cores on cluster and only use 4 of them, the 
> cluster support will not be happy.
On some clusters, it is possible to request just a part of a node (shared node 
usage). Otherwise, it may be possible to run several instances of Espresso at 
the same time on the cluster node.

Regards, Rudolf