40 Core Server is slower an 12 Core Workstation???

[dag84]

Hi everyone,

My company recently purchased a 40 Core HP Proliant DL580 server (4 processors, 10 cores each, 2.4 GHz).
This is an upgrade from our 12 core HP Z800 workstation (2 processors, 6 cores each, 3.06 GHz)

Note that we have only 19 tokens, so we max out at 24 cores.

We are seeing mixed results comparing our 12 core workstation and 24 cores on our server. Some small explicit jobs are faster, but some larger explicit jobs are slower, including example E1 from Abaqus. (

http://www.simulia.com/support/v611/v611_performance.php)

Does anyone have any idea why running 24 cores on this 40 core server is not significantly faster than a 12 core workstation?

Thanks,
Drew

 

[IceBreakerSours]

More CPUs should help in running Explicit jobs faster. I am not sure why it is not working out the way it is supposed to.

As far as I know (and have read online), there is more to it than just the number of cores. For example, a simple linear model (no large rotations, strains, linear material, no contact, etc.) may work faster on a 2 core machine than a 8 core machine. Why? Because if you force this simple model to run on multiple cores, memory I/O and data transfer speeds will start playing a role in offsetting the benefit of more cores. Also, you may wish to: a) disable hyperthreading (if it is enabled), b) not run anything except the model on the machine, c) disable all non-essential services, d) run the model immediately after re-starting the machine (in order to prevent the OS from starting unnecessary services and use available memory).

http://www.eng-tips.com/faqs.cfm?fid=376

http://www.eng-tips.com/faqs.cfm?fid=1083

 

[dag84]

Thanks for the response. We've done all items a through d and still nothing. This is happening on an Abaqus benchmarking model of a car crash, so by no means is it a simple linear model.

 

[TGS4]

Multi-core processing tend to be dominated by inter-core communication. It would appear that your inter-core pipelines are getting bogged down. Can you run some diagnostics on that? What OS are you running?

 

[IceBreakerSours]

This seems like a hardware issue but I am not a hardware geek so take this with a pinch of salt. Assuming all the components in the workstation are new, then it "may" be due to:

a) relatively poor system bus speeds,
b) low L1/L2 cache memory,
c) relatively slow memory card (with chips running at different speeds),
d) small page file (see this)
e) slower hard drive rpm, and
f) poor graphics card (only if you are running a browser, CAE, SolidWorks, etc. while running your simulation)

If your workstation is home-brewed, then, sub-optimal compatibility of components may also be slowing things down.

http://www.eng-tips.com/faqs.cfm?fid=376

http://www.eng-tips.com/faqs.cfm?fid=1083

 

[ludoLLN]

Hello everybody and especially dag84,

We have the same problem , we have bought a server with 32cores instead of 12cores before, we have 19 tokens and can so launch jobs on 24 cores maximum.
Up to now (we bought the station 1 month ago), cpu time with theses jobs on 24 cores are longer than those launched on 6 or 12cores on the old and new machine!!
We are excusively using abaqus explicite with highly non-linear (large-strain) problems.

Thanks for your help.
Ludovic.
France.

 

[TGS4]

What's the OS?