CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   Hardware suggestion for running 40M cells (http://www.cfd-online.com/Forums/hardware/130172-hardware-suggestion-running-40m-cells.html)

dv.darshan February 20, 2014 19:31

Hardware suggestion for running 40M cells
 
Hello everyone,

I have been running CFD simulation for more than 1.5 years but it had always been small cases of about 5M cells and I was running either on desktop or laptop. Currently, I am in need to perform simulation of a problem of about 40M elements, RANS calculation and to be able to finish in 1 or 2 days. Could you please suggest me the details for building the workstation? such as, what is kind of RAM amount and number of parallel cores I would be needing to perform simulations within such time limit?

Since, I am very new to this size problems I am unaware of the requirements and your suggestions will be really helpful.

ski March 5, 2014 18:44

Quote:

Originally Posted by dv.darshan (Post 475947)
Hello everyone,

I have been running CFD simulation for more than 1.5 years but it had always been small cases of about 5M cells and I was running either on desktop or laptop. Currently, I am in need to perform simulation of a problem of about 40M elements, RANS calculation and to be able to finish in 1 or 2 days. Could you please suggest me the details for building the workstation? such as, what is kind of RAM amount and number of parallel cores I would be needing to perform simulations within such time limit?

Since, I am very new to this size problems I am unaware of the requirements and your suggestions will be really helpful.

This is all very subjective, but here are some rough numbers to give you some idea. 40M cells will probably need in the region of 40GB ram. This depends a lot on how many transport equations, mesh type (polyhedral need more than structured hexa for example), etc.

Nobody can tell you how many cores you will need to finish a calculation in 2 days. It is impossible to know from the information you provide since it will depend steady/unsteady, number of iterations/time-steps required for convergence, transport equations, quality of mesh (poor quality will need more iterations on the pressure), etc. etc.

I can however tell you that 40M cells will, for most codes, scale to ~400 cores which is a cluster, not a workstation. And that will be with infiniband interconnect. If you really plan on running 40M on a workstation I think you will, for practical reasons, be limited to steady RANS at a push.

derekm April 15, 2014 10:42

from what I have read here
(cores*50E3)< cells < (Memory/1E3)


All times are GMT -4. The time now is 03:22.