Hi Lizheng, Your numbers look a little strange. First of all CPU-time is per definition the time you occupy a single processor. So if you fully occupy two processors, the CPU-time will be twice the wall clock time. Thus, your increase from 72s to 84s when going to two processors is perfectly normal and is due to communication overhead in the parallel computation. Your scaling with 3 and 4 processors seem to be very bad however.

Best regards

Leonard Lorentzen

Hi Leonard,

Thank you for your commemts.

I felt strang thing is that even for total wall clock time it also cannot be explained. For example, in my case one processor the wall clock time is 74.449s; two processors 49.705s; three processors 87.78s and four processors 669.47s. Except wall clock time for one processor, all the cpu time is almost twice times the wall clock time. what are reasons for 3, and 4 processors, the cpu time is so large?

Do you have commemts about that?

Thank you very much,

lizheng

Hello every one,

Does someone have idea why in Fluent 5.3, the CPU time for parralle calcualtion is increased with numbers of processes. I run the case in parralle calculation, I got the cpu time for one processor is 72s and 84s for two processor and 148s for three processor and 1400s for four processor.

What is reason for that?

Thank you for your help.

lizhng

Hi Lizheng,

Something is slowing down your parallel process when you employ 3 or 4 processors. Are you using a cluster of work stations or a parallel machine? Windows NT or UNIX?

Leonard

I've noticed something similar with Fluent 5.3.18 on an SGI Origin 2000. What sort of machine are you using?

Note: I re-installed Fluent 5.1.1 and this problem disappeared.

we use Sun system. Also Fluent5.3.18.

Shall we re-install this sofeware again?

Before you do this, contact Fluent support to see if there is a bug on the Sun platform.

Currently, I have two different fluent directories (fluent.old - Fluent 5.1 & fluent.new - Fluent 5.3). I copy whatever version I want to run to the Fluent.Inc directory. Make sure the license manager is shutdown before doing this. Then I compared the actual "timed" CPU runtimes for Fluent 5.1 and Fluent 5.3.

Caution: This bug may be specific to our machine, i.e. we may have messed the installation of Fluent 5.3. I don't imply that this is the case on every platform or even for the SGI in general.

If you case is not large enough, communication between computers take more time than effective calculation. I think that parallel processing is only usefull above a certain "size" of the case you want to calculate. Otherwise it is a loss of time.

What size is your mesh ? Which models do you use ?