I have ran fluent 14.5 on a cluster including 8 system with diferent number of cores and CPU frequencies (a 24 core with 3.67Ghz, a 16 core with 2.2 GHz, and 6 core with 2 Ghz)
when I read data up to 24 core,after case reading data will be read fast and CPU usage is more than 95, but when I use more than 24 cores( I test 40 and 80) data reading takes a very large time (about a day) and CPU usage fluctuates.
( I use 1G switch and cat5e cables and latency is low enough)
why? and what can I do?
regards
I'm assuming that when you use 24 cores only, all 24 will be on the same 3.67 GHz system. Where is the data stored? If it is also on that same 24 core system than you will see the best performance there. Also, do you have sufficient memory on the other systems to actually run the problem in Fluent? Switching to virtual memory is very (very) slow and will significantly degrade your performance. Also keep in mind that you will only be as fast as your slowest task.
In regards to the reading issue. If it is taking significantly longer to load the data when you use multiple system, it is like due to (a) insufficient memory and/or (b) very poor network file system performance.
Dear christofer
many thanks for your fruitful answer.
I have done tests once on 24 core and then on 3 8-core system and in both cases i see best performance, I mean that i have tested either on one system and a group of three homogeneous system and I see the best performance.
a. I used job scheduler and seems to mean that It used shared memory.
b.RAM usage percentage of systems shows that they are not fully used.(do you have any other meaning about sufficient memory?)
c.as a matter of fact I don't know what is virtual memory means and when we switch to it? may you explain it or share a link or reference about it?
d.reading task of fluent as shown in network panel seems to be done very fast ,about 3 minutes[ I mean data transfer occurs fast but this window remains and task takes a long time ] but reading data window remains. I think it seems to be about MPIs rather than sending data.
now with this description that up to 3 homogeneous system has good agreement, can network be bottleneck?
Or using standard I/O cause this problem?
regards
Virtual memory is disk space that is dedicated to the operating system to use in case the user requests more RAM than is available. It is useful to allow the system is continue under non-ideal conditions (i.e., robustness) but can be detrimental to performance. Basically, you don't want to have to use this for a CFD simulation.
So when do you see the poor scalability. The 24-core system is good. The 3 8-core systems is good. When is it poor?
Dear Christopher
many thanks for your helps
trying to use more than 24 core is poor. as an example 40 and 80 cores are very slow in reading data (and further MPI ping pongs) and fluctuation in cpu usage.
- Badges
- Science topic