CPU selection 1

[tedj]

I'm in the process of configuring a new system to run Ansys, and I'm wondering if the Pentium D Dual Core cpu is better or worse than the Xeon chips? Does Ansys take advantage of the dual core processor?

thanks, TJ

 

[TGS4]

ANSYS does handle multiple processors with ease - it was, after all, first designed for UNIX systems that almost always had multiple processors.

Just a warning though - my experience has been that the limiting factor in solution speed is NOT processor speed (or number of processors), but rather memory. I would highly recommend that you investigate a 64-bit option and then purchase as much memory as you can afford. You're money is probably better spent on memory than extra processors.

My 2-cents.

 

[tedj]

Thanks for the info. I'd like to upgrade to a 64bit system, but I still use alot of 32 bit applications. I'm also not sure if there are 64 bit drivers for all my 'old' peripherals. Or am I missing any obvious solution?

TJ

 

[TGS4]

The 32-bit apps _should_ work. 64-bit drivers for peripherals OTOH are a bit of a crap-shoot. Check out WinXP64 websites aas well as Linux distros for hardware compatibility issues. Note that ANSYS v.10 does NOT work in 64-bit mode on AMD64 XP or EMT64 XP - not yet anyways.

From my perspective, Linux is the way to go. However, this would likely throw a monkey-wrench in your old 32-bit windows apps. Definitely no easy solution.

 

[Drej]

Just a small caveat on the comment that ANSYS handles multiple processors with ease. Most of the ANSYS bulk code will not use multiple processors (MP)as far as I'm aware. ANSYS LS-DYNA of course can use MP with ease, but any general implicit solver within ANSYS cannot. There is an option to use a special solver which allows MP use, but the functionality and scope of this solver is limited, and you need to purchase a special license. This is all assuming ANSYS is running under Windows of course. Would agree heartily with the comment on memory (RAM) though - more is better without a doubt and can significantly affect solution times.


------------
See faq569-1083 for details on how to make best use of Eng-Tips.com

 

[tedj]

Thanks for the recommendations! I will forgo the expense of the dual core processor and maximize the RAM. One last question: When choosing the cpu, there are options regarding the amount of L2 Cache (1 vs 2 MB) Does the extra L2 Cache increase Ansys performance?

thanks again,

TJ

 

[TGS4]

I don't know of a direct answer to that. I would check out the benchmarks for numerical processing (good ol' number crunching) for some guidance though. Good luck!

BTW, Drej was right that some parts of ANSYS do not parallelize (for two processors - like I have) or multiplize (for more than two processors - like I wish I had) - namely the graphical boolean operations and some of the meshing. However, I can verify that the PCG solver and the Sparse solers both seem to optimize the amount of work performed on more than one processor. Since the majority of your "wait time" will be during solve, I woudl classify this as essentially well optimized for multiple processors.

 

[cbrn]

Hi,
from what I understood, most of ANSYS solvers don't "really" parallelize, i.e. they don't divide the job into "sub-systems" to be solved parallelly. It's rather Windows that tries to "balance" the load between the processors by switching the threads. Effectively, if you keep a Task Manager open when Ansys solves, you will see that it calls much more than a single thread, so the presence of several processors can be beneficial anyway.
It is a matter of fact that Linux / SCO Unix are the best platforms for computationally-intensive tasks. If I were you, I'd set up a dual-O.S., using Linux for most of the job (note that OpenOffice, among others, has a very good compatibility with MS Office; Mathematica exists in Unix version, and so does Matlab I think.
About memory, I can add some info: at my company, we recently ran a 900000 elems, >2000000 dofs, using PCG solver, on a Dual-Xeon with 3GB RAM: the CPU indicators reached 100% occupancy only for a few seconds, averagely it was 90% + 50/60%, but the available memory dropped under 4000 kB several times!
Another 2-cents: I would consider a VERY fast raid-array of SCSI discs: despite the memory availability, ANSYS writes/reads a lot... For several operations, very huge files are written, parked, and then re-read...
Level2-cache speeds-up the whole processing subsystem, so it has been benchmarked several times that going from 512kB to 1MB and then to 2MB produces noticeable effects.

Regards

 

[TGS4]

I would agree with cbrn that Linux/Unix variants are "better", but only inasmuch as they are much better at managing memory. ANSYS does in fact parallelize properly in Windows as well as Linux/Unix. This has to do with the mathematical libraries that lie beneath the complied number-crunching code. If these libraries are parellel (and there actually are different versions - serial and parallel) then the numerical library will run in parallel - and not just some Windows bastardized parallelization implementation.

W.r.t. the array of fast SCSI disks - six months ago I would have agreed with you. Now, with 16GB of RAM, I find that the only rading/writing from/to disk is with a SAVE/RESUME. It's unbelievable the speed-up that I have seen. I performed on analysis that took almost an hour when I solved it on my 32-bit XP boot. Then, when I booted into 64-bit Linux, it took 17 minutes!! Al because of the added memory-addressign capability of the 64-bit OS. Can't go wrong with that.

Cheers