Parallel distributed. ..

donminoti · March 12, 2017, 18:51

Hi
Please, I am setting up a small cluster with 4 nodes ( 8 CPU , 64 cores), plus a master node (8 cores). I want to make the head node participate to computation. To avoid batch & scripts and to launch solver from graphical interface. We'll mainly use CFX and fluent
Is it a good idea ?
Another point, is the cable lengths. The 4 nodes are linked to the switch through 2 meters cables, these nodes and switch are in separated room (lound), however the master (in my office) would be connected to that switch by a 15 meters cable!,
Could this difference in cable lengths creates a latency during solution process ?
The connection is 10GbE sfp+
Thank you
BR

Sent from my SM-G360F using CFD Online Forum mobile app

me3840 · March 14, 2017, 22:07

It's a good idea to run from batch instead of using the GUI, performance will be better and the head node will see less overhead if there's no desktop environment. Yeah, you have to learn to script a little bit, but making those scripts to automate your process will save you an incredible amount of time versus manually changing simulations all the time.

The cable lengths won't matter. The signal transmission speed is orders of magnitude smaller than the latency from the software and hardware.

Also it's a bit more difficult to install, but getting used QDR Infiniband hardware is pretty cheap nowadays, so if you don't already have the 10gige equipment, you may want to look into IB instead. 10gige is very expensive, and IB will perform better.

March 12, 2017, 18:51	Parallel distributed. ..	#1
donminoti Member belamou Join Date: Apr 2009 Posts: 40 Rep Power: 17	Hi Please, I am setting up a small cluster with 4 nodes ( 8 CPU , 64 cores), plus a master node (8 cores). I want to make the head node participate to computation. To avoid batch & scripts and to launch solver from graphical interface. We'll mainly use CFX and fluent Is it a good idea ? Another point, is the cable lengths. The 4 nodes are linked to the switch through 2 meters cables, these nodes and switch are in separated room (lound), however the master (in my office) would be connected to that switch by a 15 meters cable!, Could this difference in cable lengths creates a latency during solution process ? The connection is 10GbE sfp+ Thank you BR Sent from my SM-G360F using CFD Online Forum mobile app

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Parallel run with distributed data: How?	eskila	OpenFOAM Pre-Processing	8	September 27, 2021 06:44
Explicitly filtered LES	saeedi	Main CFD Forum	16	October 14, 2015 11:58
simpleFoam parallel	AndrewMortimer	OpenFOAM Running, Solving & CFD	12	August 7, 2015 18:45
Running CFX parallel distributed Under linux system with loadleveler queuing system	ahmadbakri	CFX	1	December 21, 2014 04:19
HP MPI warning...Distributed parallel processing	Peter	CFX	10	May 14, 2011 06:17

March 14, 2017, 22:07		#2
me3840 Senior Member Join Date: Nov 2010 Location: USA Posts: 1,232 Rep Power: 24	It's a good idea to run from batch instead of using the GUI, performance will be better and the head node will see less overhead if there's no desktop environment. Yeah, you have to learn to script a little bit, but making those scripts to automate your process will save you an incredible amount of time versus manually changing simulations all the time. The cable lengths won't matter. The signal transmission speed is orders of magnitude smaller than the latency from the software and hardware. Also it's a bit more difficult to install, but getting used QDR Infiniband hardware is pretty cheap nowadays, so if you don't already have the 10gige equipment, you may want to look into IB instead. 10gige is very expensive, and IB will perform better.