R2's 3D coupled flow solver

[p.wozniacki]

I've only now installed MPI Revision 2, and here are my initial observations re: Coupled 3D Flow Solver multi-threading:

1. While the 3D Warp solver can really take full advantage of my Intel Quad CPU (running at 100% of all 4 cores), the Flow (Fill and Pack) analyses - while claiming to be using 4 cores as well - hardly ever tax the CPU at higher than 45%

2. With the Fill+Pack Reactive Molding analysis, even though I set the solver to use 4 cores, it clearly doesn't - the log says "using 1 thread"

Why is that? While I guess the new flow solver needs some optimization tuning, and hope it'll be fater in future releases, I wonder why it doesn't use more than 1 core for Reactive Molding.

Thanks for comments,

Piotr

[raalteh]

Hi Piotr,

> 1. While the 3D Warp solver can really take full advantage of my Intel Quad CPU (running at 100% of all 4 cores), the Flow (Fill and Pack)
>analyses - while claiming to be using 4 cores as well - hardly ever tax the CPU at higher than 45%
A very good observation. The utilization is case dependant, and is also dependant on your hardware.
A possible reason for this is the speed of the Front Size Bus (FSB) or HyperTransport speed of your chip. With increasing number of threads, the speed of the analyses is to a great extent determined by the speed at which data from memory can be pushed to the analysis cores. The faster information can be pushed from RAM to the computational core (the faster the FSB or HyperTransport), the less time the CPU is 'waiting' and the more time it is actually 'working'. The more cores use the same FSB, the bigger the FSB bandwidth is a problem, the more time the multiple cores will spend 'waiting'.
The good news is that both Intel and AMD have made strides on this. The new Intel Core 7i has a few new tricks to increase the bandwidth for getting information from RAM to the cores.


>2. With the Fill+Pack Reactive Molding analysis, even though I set the solver to use 4 cores, it clearly doesn't - the log says "using 1 thread"
The release notes state that the multi-threading option is not supported for the 3D reactive molding solver. For technical reasons, the option could not be removed in the user interface, but it is documented in the help.

I hope this helps.

[AartNAAKTGEBOREN4319]

Hi Piotr and Hanno,

The new Flow (Fill+Pack) multi thread solver CAN reach high CPU usage levels.

The point is that in very large 3D filling calculations the CPU usage is 2 or 3 CPU's at the first part of the filling.
When more elements are involved i e 70 - 80 % of the part is filled, than really almost ALL of my 8 CPU's are used.
I could reach 85 to 89% CPU usage of my 8 CPU's, that is at least 7 out of 8.

The total time saved using all cpu's or 1 cpu was 1,78 times shorter in elapsed time (not CPU time ! ). (both 2010R2-SP1)

Compared to MPI 2010 R1-SP1 the MPI2010R2-SP1 version was 2,37x upto 3,27x faster running duplicate calculations on
medium to large 3D fill-pack calculations.
There was no difference in WARP calculations, because I already used 8 CPU's in MPI2010R1-SP1

The workstalion was a DELL T7400 with 8 CPU's and 32 GB memory and 4 raid 0 15k rpm disks and XP64.

Aart Naaktgeboren
Wavin T&I The Netherlands

[p.wozniacki]

Apart from the number of nodes/elements "involved", I'd still stand by my viewpoint that it may be the solver optimization matter (hopefully to be resolved soon).

A simple analogy: I'm a serious video hobbiest, and while rendering a HD video in one NLE can only tax my Quad at some 50%, the same video is being rendered at full 100% in another NLE. The same clip file, the same FSB... Yet tthe NLE that taxes my CPU fully, only uses half the time needed by the other NLE to render the same material.

Anyway, introducing the 3D Flow solver capable of multi-threading is certainly a step in the right direction!

Edited by: p.wozniacki on Aug 30, 2009 3:37 PM Edited by: p.wozniacki on Aug 30, 2009 4:27 PM

[raalteh]

One more thing to note is that using hyper threading, is not expect to provide the same speed up as real physical cores.