Unfortunately, software's codes are not written in their best way(or I think so). When you try to change from serial to 2-core parallel processor, efficiency is more enhanced than changing from 2-core to 4-core. I mean the computation capacity-core slope is getting lower while the number of cores increased. In my point of view, it shows that the relation of variables may have a better possible arrangement.
There is always a certain amount of overhead because of communications, so scaling is never perfect. With most codes you're fortunate if you get an 80 percent scaling efficiency.
On top of that, you have to keep in mind the architecture of the system. If you have two cores on a single board, communications between them will be very fast, but if your 4-core processor actually consists of two 2-core sub-sets, you suddenly get a slower communication, because now you have to move data between two separate boards. It will also change the memory loading, because the two boards don't share the same memory chip. This gets even more complicated if the code is writing parallel output and all cores have to address a hard-drive
Finally, you have to look at the simulation itself. If the physical domain is relatively small (in number of grid points) scaling will go down quickly. The optimal number of cores for a given problem depends on the code. It's very difficult to predict, so you'll have to experiment to find what works best for you.
You can try running FLUENT using journal file in background mode on parallel platform. When graphics are not loaded, the computational efficiency increases significantly.
@ Allard Jan Van Marle
many thanks to you.your scientific notes are useful for wide range of CFD inchoate students like me.
@subhasish Mitra
thank you so much.
To be honest I couldn't digest you fruitful note completely. you mean I may change reporting interval or you mean I need to do any other graphical thrick?
I've spent a lot of time and done a lot of testing on parallelization of fluid flow codes on multiple Intel processors. If you only have 2 processors, the parallelized version of the code will probably run slower than the non-parallelized version of the same code using only one processor, because there is just so much more overhead. All you have to do is pull up Task Manager to see that both processors aren't being fully utilized. With less than 4 processors it's a waste of time and effort, besides, where did you get a machine with less than 4 processors? Get some friend in the U.S. to go to Sams or Best Buy late Thursday night or early Friday morning when the doors open for the "Black Friday" sale and buy a laptop with 8 processors for $750, use it for a week, scuff it up, and then mail it to you as a "used" computer.
If you have at least 4 or 8 processors and all of them aren't running at more than 25% utilization, that's a problem with the coding that you probably can't fix. It could be a domain discretization issue, but not likely. You'll never get them to all be running at anywhere near 100% utilization.
@ Dear Dudley Benten
many thanks for your assiduous guidance.my unknown points fully obviated.
Mahdi Izadi and Rajesh Alayil, here's how to run Fluent in batch mode: https://www.sharcnet.ca/Software/Fluent14/help/flu_ug/flu_ug_BatchExecution.html
Dear Ahmad
many thanks
my fluent version is 14.5 and seems to be OK.
my problem has obviated.
regards
- Badges
- Science topic