Please wait...
×

Error

Posted Thu, 20 Aug 2009 09:41:22 GMT by Gemma
Hi,

Just wondering if anyone can explain how the parallelization of FEFLOW works?  Does this mean that even if I am just running one model, it will be quicker to run it on a multi-core machine?  If so, then would a quad core be more efficient than a dual core, or are there only ever 2 processes running simultaneously?  Thanks!
???
Gemma
Posted Thu, 20 Aug 2009 09:44:14 GMT by Denim Umeshkumar Anajwala
By default, FEFLOW runs as many threads simultaneously as cores are found. So a quad core would be more efficient than a dual core. However, actual speedup highly depends on the type of model and the time needed for different parts of the simulation. For example, parallelizing the matrix setup is more efficient than parallelizing the equation solver (only SAMG works in parallel, anyway). Output on screen is currently not parallelized at all.
Posted Thu, 20 Aug 2009 11:52:59 GMT by Gemma
Peter,

Thanks for that!  I'm also wondering if there are any general rules of thumb on how the various properties of my computer such as memory, RAM, processor etc. will affect the speed of a model run.  Does that depend too much on the particular FEFLOW model?  I'm just trying to figure out which of our company PCs will be the fastest for running regional, 3D variably-saturated FEFLOW models.

Cheers,

Gemma
Posted Fri, 21 Aug 2009 07:05:29 GMT by Denim Umeshkumar Anajwala
Gemma,

memory will only affect the run time if not enough memory is available. Then the system will start swapping to the hard disk and the speed will reduce extremely. You can check for the approximate memory demand for the simulation using Info - Storage demand in FEFLOW.
Processor clock speed has an almost linear influence, i.e., the faster the processors, the faster the simulation. When writing results files for models with many but very quickly computed time steps also the disk speed may have an influence. For larger models this is usually negligible. To improve performance, I recommend to not render the results on screen for each time step as the visual output will slow down the simulation, too.
Posted Fri, 21 Aug 2009 08:00:18 GMT by Gemma
Peter,

Thanks for that information.  It will be very useful in trying to work out which machines will be fastest for running FEFLOW.

Gemma
Posted Thu, 08 Oct 2009 20:42:04 GMT by mnovotny
Also Gemma,
CPUs with various levels of cache are usually much faster.  For example an older 2.5GHz processor with little or no cache may be slower than a new 1.5GHz processor with 3 levels of cache.
The best way to know a computers performance is to benchmark it - run the same model on each and clock it.
Milos
SWS-Denver
Posted Tue, 13 Oct 2009 16:56:30 GMT by Gemma
Thanks Milos.  The benchmarking thing is a good idea...if we ever get round to it!

You must be signed in to post in this forum.