Lower calculation in parallel than in serial

FlyingCat · January 27, 2017, 06:29

Hi everybody,

I have a strange problem. I have a case with about 300K cells. I am trying to run it in parallel. It works but it is lower than with only one core.

I have looking for solution in this forum but there is nothing. I run it on a cluster SGI UV 2000. The processors are Sandy Bridge E5-4640 @ 2.40GHz.

Even with 2 processors, it is lower (the executation time and the clock time): 1 processor with 4gb is faster than 2 with 4gb.
The problem does not come from the communication between the processors (I think) because even the execution time is bigger.

I tried to run other cases with less cells or more cells, and it is always the same result.

I also tried to run these cases on my computer, not the cluster, and parallel performance are better than serial performance with one core.

I modified the solvers, the decomposition method but even if it can improve the time of calculation, it is still lower.

If someone has an idea, I would be grateful.

FlyingCat · February 1, 2017, 10:23

No idea ?

khedar · February 1, 2017, 22:07

How did you check the communication times? And what did you change in the solver to make it slightly faster?

January 27, 2017, 06:29	Lower calculation in parallel than in serial	#1
FlyingCat Member Jean bon Join Date: Sep 2016 Posts: 43 Rep Power: 10	Hi everybody, I have a strange problem. I have a case with about 300K cells. I am trying to run it in parallel. It works but it is lower than with only one core. I have looking for solution in this forum but there is nothing. I run it on a cluster SGI UV 2000. The processors are Sandy Bridge E5-4640 @ 2.40GHz. Even with 2 processors, it is lower (the executation time and the clock time): 1 processor with 4gb is faster than 2 with 4gb. The problem does not come from the communication between the processors (I think) because even the execution time is bigger. I tried to run other cases with less cells or more cells, and it is always the same result. I also tried to run these cases on my computer, not the cluster, and parallel performance are better than serial performance with one core. I modified the solvers, the decomposition method but even if it can improve the time of calculation, it is still lower. If someone has an idea, I would be grateful.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
parallel run is slower than serial run (pimpleFoam) !!!	mechy	OpenFOAM	18	August 17, 2016 18:19
Dynamic mesh adaption in parallel calculation	xh110120	FLUENT	1	March 12, 2016 09:05
Parallel vs Serial Run-Solver manager	rohit_turbo	CFX	5	March 4, 2016 09:34
Dynamic mesh crashes in parallel, but works in serial	86lolo	FLUENT	7	June 24, 2014 06:24
Parallel run diverges, serial does not	SammyB	OpenFOAM Running, Solving & CFD	1	May 10, 2009 04:28

February 1, 2017, 10:23		#2
FlyingCat Member Jean bon Join Date: Sep 2016 Posts: 43 Rep Power: 10	No idea ?

February 1, 2017, 22:07		#3
khedar Senior Member khedar Join Date: Oct 2016 Posts: 111 Rep Power: 10	How did you check the communication times? And what did you change in the solver to make it slightly faster?