CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   FLUENT (http://www.cfd-online.com/Forums/fluent/)
-   -   IBM SP (http://www.cfd-online.com/Forums/fluent/27609-ibm-sp.html)

Johan Larsson April 5, 2000 04:55

IBM SP
 
Does anyone have any experience on running Fluent on IBM SP-machines? Are there any specific problems, is the performance satisfactory?

Volker Pawlik April 6, 2000 03:40

Re: IBM SP / communication librayr vmpi
 
One year ago I had some experience with Fluent on a SP-Cluster (256 Cpus) located at the Rechenzentrum ("center for calculations") of the university of Karlsruhe, Germany. (see e.g. http://www.uni-karlsruhe.de/Uni/RZ/). Debis system house offers cpu-time on that cluster for industrial customers. Unfortunately the hardware is not really up to date. At first the results said that using more than 2 cpu made no sense. Later on we increased the reporting interval, which, even with a single processor, augments the performance. With a 61620 cells test-case (2eq. turbulent, isothermal) we observed a speedup of 1,75 form 1 to 2 cpus, rep interval=1 and 2,05 (2,66, 3,13) increasing the cpus from 2 to 4 (6, 8) with rep interval=5. This seems to be ok. (With a lower rep-interval the speedup decreases enormously)

Earlier version of fluent5 (UNS4.?) knew a special SP2 option for the parallel solver. I don't know why it is abandoned and what the advantage was. Maybe another communication library was used.

Conclusion: no special extraordinary problems. No extraordinary speedup compared with "normal" clustering of IBM-Workstations.

That leads me to question I: Do you know s.th. about the vmpi (vendor mpi) libraries for IBM? The users's guide says that the vmpi shall perform better than the standard ones. Unfortunately it is very difficult to get information on that vmpi for IBM (where and what must be installed, etc.) and definitely I cannot use it on my machine.



All times are GMT -4. The time now is 06:19.