Page 1 of 1

Workstation Hardware Requirements

Posted: 28 Jul 2014, 18:10
by MT_Keg
Hello all,

What are the recommended system requirements for a workstation running Elmer (i.e. 4-Core with Hyperthreading, 4 memory channels, 2GB video card)?

I plan on procuring a system but am in a bind between two processors (i7-4790K and i7-4820K). The i7-4790K only allows 2 memory channels (DDR3-1600) while the i7-4820K allows for for 4 memory channels (DDR3-1866). Additionally, the motherboard for the i7-4820K (4 core/8 thread) will allow for a CPU upgrade to at 6 core/12 thread processor.

Thanks in advance!

MT

Re: Workstation Hardware Requirements

Posted: 28 Jul 2014, 21:44
by raback
Hi

Currently large parts of Elmer code cannot use more than one thread. Things are gradually changing but today I would not put too much emphasis on the multithreading. The best parallel performance is obtained by MPI. There the memory channels do have some impact but I would not guess that the difference is that significant. If you have a lot of memory it may be better to run many individiual cases rather than one parallel case. Sometimes you could have superlinerar parallel performance if you get lucky with the use of cache memory. So I would generally go for good single-core performance and sufficient memory per core.

-Peter

Re: Workstation Hardware Requirements

Posted: 03 Aug 2014, 10:48
by ChrisHarding
Hello,

Does MPI = Message Passing Interface?

Also, I had the same question and was searching today. I suspected the requirements would be similar to gaming computers.

I came across a COMSOL page that made the following recommendations. I do not mention COMSOL for advertisement reasons, but mention COMSOL because I obtained the info from their website.

What do you think about these suggestions, and can you add anything to them?

"A 64-bit operating system is highly recommended.
For a given hardware, the choice between Windows, Linux, or Mac OS X will not significantly affect performance.
At least 4-8 GB physical RAM per core of the computer is recommended.
Dual-socket or if possible four-socket nodes are recommended.
The number of CPU memory channels should be as high as possible.
A CPU with as fast a memory bus as possible is beneficial. The numbers are often measured in GigaTransfers/second (GT/s) and can be found on manufacturers' webpages (see example below).
As high memory bandwidth as possible is beneficial."

I am asking because I want to build an effective computer for doing multi-physics modeling. I am willing to spend some money in this endeavor.

Thanks!

Re: Workstation Hardware Requirements

Posted: 04 Aug 2014, 00:05
by raback
Hi

The specs for COMSOL are probably quite good for Elmer too.

Elmer excels in MPI scaling often better than COMSOL. In Windows the whole MPI business is more problematic and hence a Linux system is advisable. However, recently there has been also MPI versions created for Windows but this has had a rather low priority.

-Peter

Re: Workstation Hardware Requirements

Posted: 04 Aug 2014, 10:21
by ChrisHarding
Hello,

Question: When thinking about multiple core processors, will Elmer Multi-physics take advantage of processor multitasking? I assume it will since you mentioned MPI[1]. Since I am learning, I thought I would make sure before I concentrate on such processors as Intel Xenon.

I was reviewing some processors today, looking at the benchmarks, and it seems that Intel has better performance than AMD. Some research has seemed to indicate Intel is a better performer as well[2].

Since I don't think I can afford a cluster or US $9,000.00 dollar workstation computer with dual processors, I really need to try and optimize my system. As such, I am paying attention to number of processor cores, higher cache RAM, and faster system bus. If I build my own instead of buying from Dell or HP, I will probably get an dual processor capable board so I can add on in the future.

Many such boards have significant GPU capabilities. I know Ansys is starting to take advantage of GPU's, but COMSOL has said the mount of RAM on current GPUs is not enough to support their solvers. Still, I assume nVidia, etc, might eventually develop a work around that would make CUDA more beneficial. In some cases, the performance of GPU over CPUs is quite remarkable[2].

Question: What are your opinions on the above Dr. Raback?

Of course, I suppose that today's motherboards might not be able to handle future GPUs though. I just hope for an opportunity to scale up over time. I am learning how to use multi-physics, learning about turbulence models, I make mistakes, I currently have a slow computer, and it can take a while before I discover non-convergence, etc. As such, my progress is a bit slow. I cannot afford or manage a cluster, but want a good workstation.

Since this topic/thread might be visited in the future, I will share some of my future references. :)

References:

[1] Kuron, Mike. HPC Best Practices for Structural Mechanics—Part I. caeai.com[online]. 2014. Available from: https://caeai.com/resources/hpc-best-pr ... ics-part-i

[2] Thibault, Julien; Senocak, Inanc. CUDA Implementation of A Navier-Stokes Solver on Multi-GPU Desktop Platforms for Incompressible Flows, January 2009. scholar works.boisestate.edu[online]. 2014. Available from: http://scholarworks.boisestate.edu/cgi/ ... ng_facpubs

Re: Workstation Hardware Requirements

Posted: 04 Aug 2014, 12:16
by raback
Hi Chris

There are no plans to utilize GPUs for Elmer. The problem is that this would require rewriting of large chunks and the efficiency of GPUs for sparse linear systems is not that great. We have been looking for Intel MICs because the tricks there will allow to utilize also standard multicore processors more efficiently, see
http://www.cerfacs.fr/files/cerfacs_alg ... ckling.pdf

I would go for Intel multicore CPU. The option for dual CPU may be good, but the option for GPUs has no relevance.

-Peter

Linux Distribution and Comiling From Source

Posted: 05 Sep 2014, 02:01
by ChrisHarding
Hello,

Well, I have my new workstation! Before I start and get lost in multi-physics, I want to get FreeCAD on my workstation so that I can build nice geometric structures to be meshed. I am completely new to Linux though, I decided on CentOS, and I must compile my own FreeCAD because no FreeCAD RPMs exist for CentOS. There are some for Fedora, Ubuntu, OpenSUS, but not CentOS.

Dr. Raback suggested Linux because of MPI, and I chose CentOS because COMSOL validated Red Hat Enterprise, Intel validated Red Hat Enterprise, etc. and CentOS is Red Hat code. I have also read cfd-online posts where users had some compatibility troubles with, as an example, Ubuntu.

Still, I thought I would ask you all if you prefer one Linux distribution over another. Also, I thought I would ask if you can suggest any good literature that would teach me how to compile and create packages. I have found the following so far[1-5].

I was also told CentOS might be similar to Fedora, but I have read that they vary as well. As mentioned FreeCAD has a compile example for Fedora 20.

Finally, I assume Elmer Multi-physics supplies a trusted and compatible package for CentOS. Is that a good assumption?

References:

[1] CentOS. Software Installation from Source. Retrieved (2014, Sep. 04). wiki.centos.org[online]. Available from: http://wiki.centos.org/PackageManagement/SourceInstalls

[2] DigitalOcean. (2013, Aug. 30). How To Compile and Install Packages From Source Using Make on a VPS. Retrieved (2014, Aug. 30). digital ocean.com[online]. Available from: https://www.digitalocean.com/community/ ... e-on-a-vps

[3] CentOS. Set Up an RPM Build Environment under CentOS. Retrieved (2014, Aug. 30). wiki.centos.org[online]. Available from: http://wiki.centos.org/HowTos/SetupRpmBuildEnvironment
-Never build RPM as a root; should use an unprivileged user.

[4] CentOS. Part II. Package Management. Retrieved (2014, Sep. 02). centos.org[online]. Available from: https://www.centos.org/docs/5/html/5.1/ ... ement.html

[5] RPM. (2000). Maximum RPM: Taking the Red Hat Package Manager to the Limit. Retrieved (2014, Aug. 31). rpm.org[online]. Available from: http://www.rpm.org/max-rpm/

Re: Workstation Hardware Requirements

Posted: 08 Sep 2014, 11:17
by mzenker
Hi,

you could check out CAELinux (http://www.caelinux.com) which contains FreeCAD, Elmer, Salomé and more FEM-relevant tools. It is Ubuntu-based.
I didn't try it myself - although I would love to do so, I don't use Elmer under Linux myself (no Linux desktop computers in the company I work for).
Since the Elmer team works under Linux, they might give you some additional hints and/or recommendations...

HTH

Re: Workstation Hardware Requirements

Posted: 08 Sep 2014, 19:46
by ChrisHarding
Thanks mzenker,

Yeah, I decided to go with Kubuntu because of FreeCAD forum advice about the attractiveness of "KDE" desktop environment and because CAELinux used Ubuntu.

Personally, I don't want to use a LiveDVD distribution though because of reduced performance. I do plan on installing the same software with a few extra additions like open source DWSIM for processes and optimization.

Have a good day!

Re: Workstation Hardware Requirements

Posted: 12 Sep 2014, 12:10
by mzenker
ChrisHarding wrote:Personally, I don't want to use a LiveDVD distribution though because of reduced performance.
CAELinux can also be installed on hard disk, as they say on http://www.caelinux.com under "New release: CAELinux 2013".

Matthias