Dual-core CPUs and Pro/Mechanica?

[MichailS]

Hi

I'm doing a set of rather time-consuming analysies in Pro/Mechanica (Wildfire 2.0) right now, and I was wondering if there might be a benefit to use a multi-CPU computer instead?

The prices of dual-core CPUs are nice nowadays, it might even be reasonable to put together a quad Opteron if I can motivate it with saved time.

So pray tell - is there a way to enable Pro/Mechanica to utilize multi-core systems? Some flag to add to the msengine batch runs?

Thanks

 

[Israr]

I would suggest go for 64-bit technology if you are using Windows Xp.


There is an option


CPUs to use please search it.


We use dual processor and the performance of Mechanica is much better.


Israr

 

[MichailS]

Cool! Thanks!

Do the dual CPUs cut the computing time in about half? Or is the gain marginal?

I'm asking because I recall that the multithreading in Pro/Engineer is rather weak (only a few operations speed up) whereas other FEA applications' (ANSYS, LS/Dyna) performances happily multiply with the number of CPUs.

We would most certainly use Windows X64.

 

[compujas]

Just to clarify, there is a difference between dual-CORE and dual-CPU. Dual core is a relatively newly released technology that has two cores in the cpu, while Dual CPU is old news and simply drops two processors on a motherboard. As far as functionality, I don't know the difference since I don't research hardware until I'm interested in buying. It just sounded like there was a bit of confusion between the two.

 

[dr_gallup]

There are significant architectural differences between the various dual core implementations, even within the same brand, mainly in the implementation of cache and memory access. So there is no hard and fast rule of thumb on performance improvements. Running your own benchmark on your own data is pretty much the only way to know for sure. In general, dual core should perform roughly the same as dual CPU and in some cases better when they share cache. But in any case, you will never get exact performance scaling with multiple processors

 

[MichailS]

I'm somewhat aware of the architectural aspect of dual-CPU systems. The dual-core/single-socket Opterons (and the Athlon64 X2) have on-die memory managers so they behave pretty much like single core/dual-socket systems and should not really be compared to Intels Xeon that has to communicate with its memory via a rather tight front-side bus past the northbridge. Quad Xeons are laughably starved on memory bandwidth.

The bandwidth of the Opteron setup is something like four time as great as that of the Xeon. Thus, in many settings where the Xeon chokes badly, the Opteron does not have as much of a problem.

The bottom line is that multiple-core/socket Opterons get much closer to their theoretical top performance than the Xeons. The performance difference between a single core Opteron and a quad Opteron _can_ be close to 4x whereas a Quad Xeon is rarely even twice as fast as a single Xeon.

Then again - the software is much more the deciding factor in that equation than the hardware - it is not at all necessarily the case that RAM is such a big bottleneck although this is commonly the case - hence I am curious whether Pro/Mechanica does utilize multiple cores (or CPUs if you will) well. Because _Pro/Engineer_ certainly don't...

If Pro/Mechanica speeds up well in dual/quad configurations, it might be the case that we will get such rigs.

 

[SlavoM]

Hi,


Another solution of problems with time-consuming analysis isdividing solving for more computers. This funcionality is supported by Pro/E W2 and Pro/M W2, I think.

 

[MichailS]

Well, we use all of our workstations so it would probably not be a benefit to buy several more computers instead of one or two dual/quad numbercrunchers.

Unless we combine the techniques and divide the solving on several dual/quad numbercrunchers...

This might be a stupid question, but how do I set the solver to use more threads? I can't find anything about it in the run settings in the design studies menu in the integrated environment. Do I have to tweak a config file somewhere or use the standalone version?

 

[ankarl]

Here are some hints on parallell processing for Mechanica:


[url]http://www.ptc.com/cs/help/proengineer/wildfire/html/usascii /proe/promec/run/structure/reference/strgy_run_parallel_proc .htm[/url]


(PTC account reqd)

 

[MichailS]

Thanks, but unfortunately I'm a consultant.

That means I have no PTC account myself, and I seriously doubt that anyone in this house will dig one up for me with less means than torture.

 

[dr_gallup]

<h3>Strategy: Running the Engine with Parallel Processing</h3>



By default, Pro/MECHANICA automatically selects the number of CPU's
it will use during a Structure or Thermal run. For systems with one CPU,
Pro/MECHANICA always uses standard serial processing. For parallel processor
machines, Pro/MECHANICA attempts to use all CPUs on the machine. In certain
situations, however, you may want to limit the number of CPUs the software
can use. When making this decision, you need to balance the benefits of
parallel processing against its limitations.



The main advantage of parallel processing is that it can improve solution
times. Parallel processing is most likely to improve a design study run
time under the following conditions:

<ul><li ="kadov-p-CTopic-Text-Bulleted">Only
a single user is using the computer at the time of the run.<li ="kadov-p">

The computer is equipped with a fast disk, ample
memory, and individual processors that are fast in and of themselves.<li ="kadov-p">

The ratio of elapsed time to CPU time is close
to a value of 1 with one CPU. <li ="kadov-p">

Your model is large, solid, and blocky in shape.[/list]

Even if these conditions are present, your elapsed
run time might improve only slightly when you run the Structure or Thermal
engine in parallel processing mode because parallel processing accelerates
just a few phases of the overall solution process.

The main disadvantage of parallel processing is that,
if you are running other jobs at the same time as Pro/MECHANICA, you can
experience performance bottlenecks. Thus, unless you expect substantial
gains from using all CPUs in your system for the engine or you have no
other jobs planned for your system during the engine run, you may want
to limit the number of CPUs dedicated to the engine.



If you do not want to use all of the CPUs in your system for the engine
job, you can set the environment variable MEC_NUM_THREADS to the anticipated
number of idle CPUs on the machine. For example, on a 4-CPU machine with
one CPU-intensive job running in addition to the parallel job, set MEC_NUM_THREADS
to 3.



Note that Pro/MECHANICA does not support parallel processing on the
HP 32-bit platform.

 

[ankarl]

What it basically says is that Mechanica will automatically determine how many CPU's you have and use them all by default.


can be manually overrided using the MEC_NUM_THREADS config option


Parallell processing is not supported on HP 32-bit systems.


My own experience here is that it doesnt matter how many CPU's you have if your RAM isnt large enough or your discs are slow.





OOPS, Doc was quicker than me (again)....

(copyright issues here .....????)
Edited by: ankarl

 

[MichailS]

Heh, silly me. That very file resides on my harddrive in the same folder.

your elapsed run time might improve only slightly when you run the Structure or Thermal engine in parallel processing mode because parallel processing accelerates just a few phases of the overall solution process.

Bummer.

Actually, we've got ANSYS in house as well, although I seem to recall the management speaking of phasing it out in favour of Pro/Mechanica and whatever Catias own solver is called.

Do you think that using the ANSYS solver in Pro/Mechanica would work better with multi-processor rigs than Pro/M:s own solver?

Thanks for the help, by the way!
Edited by: MichailS

 

[klheureux]

I just got a new system with dual core-dual processorfor running Mechanica.


I did a benchmark on a run from my old machine.


My old machine was a Dell 450, dualZeon (2.66 mhz ea), 2 gb ram, 32 bit XP


New system is a Dell 470, dual core-dual processor (2.80 mhz), 4gb ram and 64 bit XP


The new system shows 8 processors in the task manager. I had affinity set to all 8 while running just Pro-E and Mechanica.


The run on my old system took 10.87 hours to run, on the new system 3.54 hours


Let me know if you need any help.


Email me direct @ kplheureux@sbcglobal.net

 

[wskunz]

Here is the result of a Mechanica benchmark we ran at our company.


On our existing Pro/E systems, Dell Precision 370, 3.4 GHz P4, 2GB RAM, nVidia 3400, XP Pro 32bit, WF2 M050.A test analysis took 2.5 hours.


The same hardware running XP64 Pro & WF2 took 45 minutes.


On a new Dell Precision 670, 2-3.6 GHz Xeons, 6 GB RAM, nvidia 3400, XP64 Pro & WF2 took 20 minutes.


So you can see the the 64 bit software made a bigger difference than the much more powerful and expensive hardware.


We only saw a small improvement in Pro/E performance on these same systems.


Bill
Edited by: wskunz

 

[klheureux]

I prefer the extra capacity of the dual core-dual processor.


I do a lot of NC programming, tooling design and high end surface work. With the dual core I can run Mechanica while working on otherprojects.


It was only about $700 more for the Dual core system.

 

[MichailS]

The new system shows 8 processors in the task
manager. I had affinity set to all 8 while running just Pro-E and
Mechanica.


The run on my old system took 10.87 hours to run, on the new system
3.54 hours

On our existing Pro/E systems, Dell Precision 370,
3.4 GHz P4, 2GB RAM, nVidia 3400, XP Pro 32bit, WF2 M050. A test
analysis took 2.5 hours.


The same hardware running XP64 Pro & WF2 took 45 minutes.


On a new Dell Precision 670, 2-3.6 GHz Xeons, 6 GB RAM, nvidia 3400,
XP64 Pro & WF2 took 20 minutes.

Very interesting! I don't quite know what to make of these numbers, it
seems very strange indeed that merely the transition to 64-bit systems
would make such a big difference. I know that the x86-64 ISA allows for
more data registers and tricks like that, but usually it only accounts for a
speedup in the magnitude of 10-20% - they SAY. I myself have
noticed a 25-40% speedup in a benchmark I ran on two systems (one 32-
bit Linux and the other a 64-bit) but that test wasn't very scientific.

So while 64-bitness can account for a lot, I fail to see how it could speed
up computations to less than a third, unless the rig was starved for
RAM in the first setting? Boggles my mind!

klheureux: It is a widely accepted and rather sad fact that Intels
Hyperthreading is decelerating performance rather than accelerating it in
most cases, so I suggest that you disable it in the BIOS and see if you gain
another 10% or so - despite the fact that it will only look like 4 CPUs in
the Task Manager. Might make a difference to the positive, might not.

I did a quick check of the rigs the other guys use around here (while I
have a single-CPU P4 they have dual Xeons) and it seems as if Pro/
Mechanica does indeed utilize multi-CPU capabilities, but it doesn't do it
very well. The combined CPU utilization was wavering at some 50-80%.
Ideally, they should be smacked up at 100% all the time.

I'm underway of talking the management here into letting me build a
quad Opteron, so if it works out then I might be able to post a nice
performance comparison myself. =] I'll even try a few ANSYS runs to see if
it utilizes multithreading more efficient than Pro/M does. Since we are
pressed for time and have ANSYS in house, it might actually be rational to
use it in place of Pro/M.

Thanks for your inputs!

 

[Miko]

Hi all. I would like to start with study the analysis in Pro-MECHANICA but I dont know if is my computer suitable. I am working with MDO,MDX,ISDX,ARX modules and for this kind of work is this computer pretty good, but for Mechanica?


My computer:

CPU> Athlon XP (32Bit) 3200+ overclocked by watercooling of Northbridge and CPUat 2,45GHz


DDR RAM> 2x256 MB PC3200 (Dual Channel)


VGA> Ati Radeon 9800XT 128MB GDDR


Can anybody tell me if is this computer good? Thank you.


With Regards:Michal Mascenik

 

[Israr]

Miko,


you have to increase RAM to atleast 1 Gb and VGA of 256 Mb is recommended for Mechanica.


Israr

 

[Miko]

Thank you Israr..