I am running CFX Solve 11.0.1 on a 8 core / 8GB RAM system. The system is running Windows XP Pro 64bit.

If I specify 8 partitions local parallel execution, the system runs under only 50% load most of the time, going to 100% only in shorter periods during each iteration.

The mesh is <2 million elements, the total RAM use under 3GB so the system has plenty of free RAM when running.

I was wondering if anyone has some tips on how to tweak the solver to use the available resources more efficiently?

It seems to me I could get almost half the solution time if all 8 CPUs where running at 100%.

Cheers, /c

Forgot to mention it is the 64bit CFX version as well.

I'm currently trying to get licences for running CFX in parallel for the first time. So, I don't have experience of this, but I can pass on some advice I was given by ANSYS during discussions about this:

They recommended to me that dividing a model so that each processor is using about 1GB of memory is about right. They said that if you have less than 1GB of memory in use on each processor then the overheads of communication between the processors will cause a slow down. If you have more than 1GB per processor you may not get the full parallel efficiency available to you.

So, perhaps for a 3GB sized problem it is simply not efficient to try and use all 8 processors? Perhaps the best efficiency will be around 3-4 processors?

I hope others who have direct experience of parallel processing will also reply to you with their advice.

Regards, andy

(PS: I am very jealous that you have so many licences available to you!)

Hi,

If you have 8 parallel licenses then for most simulations you will get healthy speed-ups, even though the size of the model is not huge. You might as well use them!

This problem can be caused by: 1) Lots of file IO - the master node has to do the file IO as it is not parallelised. That means if you write lots of results, backup or output files you will loose parallel efficiency. 2) Unbalanced partitioning - Sometimes the partitioner does a poor job of partitioning. This usually occurs in models with lots of small domains connected by interfaces. Not much can be done here, but sometimes using a different partitioner can help. 3) Unbalanced node speeds - If the cluster you are running on has machines of varying speeds then obviously the faster machines will have to wait for the slower machines. You can offset this by modifying the partition size weighting but in my experience this rarely improves things dramatically. The important thing here is that clusters work best when all machines are the same, and there is no point in adding a slower machine to a cluster just as you have a free license as it will probably just slow the entire cluster down.

Glenn Horrocks

Thanks Glenn! Unfortunatly none of the above applies

1) IO: I use only the default monitor points, no backups or intermediate results.

2) Partitioning: The problem has one domain only, no interfaces. Don't know how to judge the job of the partitioner, I guess you can output the partitions and look at them in post but what would that tell me?

3) Node speeds: This is a single 8 CPU machine, so the hardware is by definition exactly the same.

Since the partitions have less than 0.5GB space each maybe my problem is that the model is too small That's good then because it means I can add detail without loosing solution time!

I will run some tests @ 4 partitions and report findings...

Have you tried the MPICH? It should be faster on windows compared to PVM.

No, but I tried and this fails for unknown reasons. Seems to be an installation problem. (Unfortunatly I am in no position to require the sysadms assistance, I'm only borrowing time from an infrequently used system...)

This is the console output when trying to use MIPCH: Unknown option: -d missing num_proc flag: -n, -np, -hosts, or -localonly. Unable to parse the mpiexec command arguments. ReadFile() failed, error 109 unable to start the local smpd manager. An error has occurred in cfx5solve:

The ANSYS CFX solver could not be started, or exited with return code 255: . No results file has been created.

OK, managed to get the MIPCH2 solver problem fixed, some install config.

"Dr Flow Squad" was correct, solutions now run at 100% on all 8 CPUs. It's quite a lot faster.

Thanks.