Small cluster configuration for pump simulation at CFX

Nevel · February 2, 2010, 08:36

Dear colleagues,
could you please recommend a configuration of workstation or small cluster for pump simulation at CFX?
The pump is centrifugal and single-stage with volute and diffuser.
Steady-state (frozen rotor) and transient simulations.
Combination of hexa and tetra/prism meshes.
Whole model number of cells is about 8 millions.
Required time for 1 iteration is about 100 seconds.
I located this thread at hardware forum but gave no answer, so I hope you can advice something.

Thanks a lot for your replies!

ghorrocks · February 2, 2010, 18:08

Get as many parallel licenses and computers as you can afford. The more you get the quicker it will go. If you are on a commercial license spend your money on top-of-the-range CPUs (like the intel X5570). If you are on an academic license it may be worth considering lower spec CPUs but more of them.

As for designing a system to complete simulations at a given speed - well, you will have to benchmark that yourself. But 1 iteration in 100s does not sound very fast. You should be able to achieve this.

Nevel · February 3, 2010, 09:31

Thanks Glenn.
What do you suppose about number of these processors?
Is speedup curve close to ideal one for real turbomachinary simulation with 8, 16, 32 nodes (not for only benchmarks)?

ghorrocks · February 3, 2010, 23:37

The optimum number of processors depends on factors including budget, uptime requirement and maintenance of clusters. These are things you have to judge for yourself.

A real turbomachinery simulation with several million nodes or more should scale pretty well to 8 and probably 16 processors. For 32 and higher processors you will probably need to consider more upmarket networks like Infiniband. But really you need to benchmark all this stuff yourself on your simulation.

February 2, 2010, 08:36	Small cluster configuration for pump simulation at CFX	#1
Nevel New Member Join Date: May 2009 Posts: 8 Rep Power: 17	Dear colleagues, could you please recommend a configuration of workstation or small cluster for pump simulation at CFX? The pump is centrifugal and single-stage with volute and diffuser. Steady-state (frozen rotor) and transient simulations. Combination of hexa and tetra/prism meshes. Whole model number of cells is about 8 millions. Required time for 1 iteration is about 100 seconds. I located this thread at hardware forum but gave no answer, so I hope you can advice something. Thanks a lot for your replies!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
nucleate boiling simulation in CFX	Anil	CFX	3	August 25, 2010 15:18
Numerical Simulation of Gas Metal Arc Welding By CFX	PATELAM	CFX	0	November 19, 2009 12:24
Multiple Configuration Simulation	Tristan	CFX	0	November 14, 2009 00:01
Running CFX on a cluster	jvk	CFX	9	September 19, 2002 23:22
Fire simulation with CFX 5	Jens	CFX	2	March 4, 2002 06:44

February 2, 2010, 18:08		#2
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,872 Rep Power: 144	Get as many parallel licenses and computers as you can afford. The more you get the quicker it will go. If you are on a commercial license spend your money on top-of-the-range CPUs (like the intel X5570). If you are on an academic license it may be worth considering lower spec CPUs but more of them. As for designing a system to complete simulations at a given speed - well, you will have to benchmark that yourself. But 1 iteration in 100s does not sound very fast. You should be able to achieve this.

February 3, 2010, 09:31		#3
Nevel New Member Join Date: May 2009 Posts: 8 Rep Power: 17	Thanks Glenn. What do you suppose about number of these processors? Is speedup curve close to ideal one for real turbomachinary simulation with 8, 16, 32 nodes (not for only benchmarks)?

February 3, 2010, 23:37		#4
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,872 Rep Power: 144	The optimum number of processors depends on factors including budget, uptime requirement and maintenance of clusters. These are things you have to judge for yourself. A real turbomachinery simulation with several million nodes or more should scale pretty well to 8 and probably 16 processors. For 32 and higher processors you will probably need to consider more upmarket networks like Infiniband. But really you need to benchmark all this stuff yourself on your simulation.