Continue to Site

Eng-Tips is the largest engineering community on the Internet

Intelligent Work Forums for Engineering Professionals

Register Log in
  • Congratulations The Obturator on being selected by the Eng-Tips community for having the most helpful posts in the forums last week. Way to Go!

Workstation PC Build for Large Deformation Finite Element Analysis (Explicit)

Status
Not open for further replies.
lyonisk888

lyonisk888

Civil/Environmental
Aug 7, 2015
8
Hi Everyone,

I am currently thinking of building a new workstation PC for ABAQUS Finite Element Analysis, I am running ABAQUS CAE 6.12 on my computer with these specifications:

CPU: Intel Core i5-4690K 3.5GHz Quad-Core Processor 6MB cache
GPU: Nvidia GTX 970 4GB GDDR5
SSD: Samsung 850 250GB 2x (Raid 0)
HD: Seagate Barracuda 1TB
RAM: Corsair Vengeance 1600MHz DDR3 CL9 16GB

The simulation is taking days and I am wondering if it is it worth building a workstation PC with 32GB of DDR4 memory and an Intel Core i7-5820K 6 core processor with 15MB cache to reduce computational time? I am doing a Large Deformation Finite Element Analysis (Explicit) using the Coupled Eulerian Langragian(CEL) method. Feedback from others who have done FE analysis (Explicit) using abaqus on workstations vs servers or desktops would be great!

Thanks For taking your time to read this!
 
Replies continue below

Recommended for you

Sort by date Sort by votes
Unfortunately, you may not have much choice; CEL is a computationally expensive analysis. Running a 'simple' CEL analysis on tens of CPUs on a Linux cluster can easily take hours. That said, all is not lost. You can try to coarsen the Eulerian mesh as a first step; that will give you the biggest bang. Another option is to try *Eulerian Mesh Motion (see the manual for details).

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
If its a one and done job, you could try contacting a company that provides "cluster for hire" services (and supports Abaqus) for a quote? Here is one I found on google:
If you are in academia there are often state-funded HPC providers that will support your research free of charge (i.e. give you access to a cluster).

Otherwise I would exploit every possible method of reducing your analysis time and stable time increment before committing to the investment.
 
Upvote 0 Downvote
In addition to Dave's suggestions, you might want to look into increasing the density of the materials because mass scaling won't work in CEL (as long as the physics allows it). But, to be sure, you should request output to ensure the averaged density (DENSITYAVG) makes sense. Also, keep ODB frames and output to a minimum (e.g., LE, PEEQAVG, U, etc. may be avoided).

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
Dave & IceBreaker Thank You for your advice, it has taken two days for the analysis to run for just 1 second of time step -.-
Coarsening the Eulerian mesh is not really an option as I will have to do a mesh convergence analysis. I don't have access to HPC in my University so I think I should probably just let my computer run constantly and hope for the best as it seems like there is not that much I can do and Hiring a cluster seems a little out of my personal budget for my thesis.
 
Upvote 0 Downvote
You can do the mesh convergence analyses once you have ensured the analysis is doing what it is supposed to be doing. Your first analysis need not be one with a very fine mesh.

By the way, one second is a LOT of time for an explicit scheme (which is what CEL uses), particularly with the kinds of element sizes and numbers involved in CEL analyses. Half a second or less are typical times for such analyses.

In addition to previously mentioned workarounds, you might also wish to look in to BCs such as non-reflecting ones for a smaller Eulerian domain. Also, you can be clever about your mesh seeding so you have a fine mesh only in areas of high interest.

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
To minimise the Eulerian domain I used a quarter model and applied a symmetrical/encastre boundary condition as the primary analysis is only in one direction. I did a couple of analyses using large elements but they gave some misleading results but now that I have partitioned the eulerian element and used dense mesh in areas of high interest. I had no idea how computationally intensive CEL and finite element analysis is. Would having more CPU cores and memory significantly help or not really for when i need to run mesh convergence analyses?
 
Upvote 0 Downvote
Your university may have access to a state/national HPC resources, even if there isn't one at your university. I'd be very surprised if there is no way for you to gain access to a cluster.

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
I found out that the only supercomputer/HPC centre in my area is called Pawsey here in Perth, Western Australia. The problem is I would have to submit an application which would only be assessed at the end of December by then my thesis would already be due -.-. I was told computational time would be long, no one told me days long I was naive to think that it would take a couple of hours to simulate a simple anchor penetrating the soil I guess...
 
Upvote 0 Downvote
Unfortunately, some of the advanced analyses like CEL or SPH etc. aren't truly ready for widespread usage. They are simply too expensive. Can you not, somehow, perform a Lagrangian analysis?

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
The large deformation compromises the mesh structure and gives some really misleading results. I think I might just build a workstation of anyways because I used up 80GB just for a step time of 1 second and I have to do at least 5 seconds....
 
Upvote 0 Downvote
By the way, mesh convergence for CEL also includes Eulerian domain convergence (which will further increase the computational cost), wherein you increase the size of the Eulerian domain to check what the reflections from the boundaries are doing to your solution.

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
I'd set the job running last thing on a Friday to run over the weekend, and just hope that when you arrive back on Monday morning you don't see a message on the screen that says "Previous job name exists, do you wish to override (Y?N)...". It happens.

 
Upvote 0 Downvote
Yea i'm not looking forward to this computational time its going to kill me!
 
Upvote 0 Downvote
One thing I learned by watching an FE 'guru' is that you should play a LOT with a coarse mesh model so you can make lots of mistakes and fix them quickly, and finalize on a set-up. Only then should you make the model 'realistic' and do your actual analyses. Otherwise, along corus' line of thought, you end up spending/wasting hours/days on inaccurate/imprecise/junk models. This is even more so true for expensive analyses such as CEL.

By the way, check out the Director's Allocation Scheme at Pawsey. It may not be much but it may very well be an order of magnitude or two better than your set-up.

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
They use some amazing haswell Xeon processors that are several magnitudes higher than I would ever afford... But you do have a good point in terms of using coarse mesh and making heaps of mistakes first before doing your final analysis!
 
Upvote 0 Downvote
Do you need to model the soil? what are you trying to measure?

If your computational time is going to limit the amount of useful work that you can do, you should consider simplifying your analysis. If you are only interested in the behavior of the anchor, for example, could you make some sort of simplifying assumptions with regards to the loading and run a static analysis instead?

Dave
 
Upvote 0 Downvote
Hi Dave! Thanks for your reply! Unfortunately I am analysing the soil behaviour as the anchor penetrates, settles and eventually pull out. I am trying to measure embedment depth and pullout capacity. Is there any suggestions to ease this situation?
 
Upvote 0 Downvote
While I do not know, odds are there are some relevant examples in the documentation. I know for sure that pore-pressure elements are typically used in computational geomechanics.

*********************************************************
Are you new to this forum? If so, please read these FAQs:

 
Upvote 0 Downvote
Status
Not open for further replies.

Similar threads

S
  • Question
ABAQUS Dynamic Explicit: Stent Crimping
Replies
12
Views
4K
IceBreakerSours
IceBreakerSours
DavidMcArthur
  • Locked
  • Question
ABAQUS Explicit Damage
Replies
2
Views
499
DavidMcArthur
DavidMcArthur
Maria Kxr
  • Locked
  • Question
Abaqus Explicit - Error in restart analysis
Replies
5
Views
351
FEA way
FEA way
S
  • Question
facing error in Abaqus explicit simulation
Replies
3
Views
741
FEA way
FEA way
C
  • Question
Generalised plane strain using Abaqus Explicit
Replies
1
Views
796
FEA way
FEA way

Part and Inventory Search

Sponsor

Back
Top Bottom