CFD Online Logo CFD Online URL
www.cfd-online.com
[Sponsors]
Home > Forums > General Forums > Hardware

Pre- and Post workstation - need inputs

Register Blogs Community New Posts Updated Threads Search

Like Tree10Likes
  • 2 Post By abdul099
  • 1 Post By larsenmm
  • 1 Post By scipy
  • 1 Post By scipy
  • 1 Post By abdul099
  • 1 Post By bindesboll
  • 1 Post By bindesboll
  • 1 Post By JBeilke
  • 1 Post By bindesboll

Reply
 
LinkBack Thread Tools Search this Thread Display Modes
Old   January 5, 2013, 08:27
Post Pre- and Post workstation - need inputs
  #1
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Hi all

I am searching for inputs on how to build a strong workstation for Pre and Post-processing. Below is described the software/hardware used, how I work and what I work on and lastly what hardware I have been considering.

All the solving is going to be running on a small 32 core cluster (Bought at the same time as this workstation). Software is ANSYS and SpaceClaim. More specifically I clean-up the CAD's in SpaceClaim (imported from Solid works), import it to workbench, Mesh using ANSYS Meshing, setup the simulation in CFX-Pre, solve using the ANSYS Solver Manager (on cluster) and Post process using CFD-Post. The cases as far as i know are located locally on the workstation when Pre-and Post-processing.

The hardware I use currently is primarily a HP Probook 6360b laptop, i5 2520, 16Gb of memory, on-board graphics, 120 Gb SSD, and my simulations are running(solved) on an old HP Z600 with two Xeons E5520, 12 GB of memory, a ATI FirePro V7750 (FireGL) and one 2 TB SATA hard drive. As the laptop is faster for Pre and post, I use that most of the time.

The cases I run is multiphase simulations of Diesel after treatment systems where focus at the moment is on optimisation of the Adblue injection upstream of the SCR catalyst. This includes multiple phases, water evaporation, chemical reactions and off course particle tracking. Due to the complex CAD geometries I use tet cells for all the complex parts. Else I try to have as much quad-mesh as possible. Mesh counts on the cases i run at the moment are in the range of 3-9 mio cells.

I often find myself sitting and waiting for the computer to finish loading, meshing, Post processing which is quite irritating. To put numbers on, loading a typical CAD from SpaceClaim to ANSYS Meshing takes 1-4 min, Meshing takes from 5-60 mins, loading a already processed file in CFD-Post takes 3-15 mins. It all adds up and it slows my work down, I try to switch between several tasks/assignments to utilise my waiting time as good as possible and that is not ideal. In general it is the waiting time and the bad graphical performance that bothers me the most.

As a side note, I am currently limited to smaller cases and lower mesh counts due to the hardware i currently work on. As soon as the cluster and my new workstation is up and running, case sizes can be expected to increase somewhat.

At the moment I am limited to HP as the supplier, as that is what our company uses. I have a meeting with out IT department in the middle of January, to discuss or tell them what I want. Budget: Need as much Bang-for-the-buck

This is what I have been looking at until now:

CPU: Xeon E3-1290v2, Quad core, 3.7 Ghz (turbo speed up to 4.1) or an i7-3770, Quad core, 3.4 Ghz (up to 3.9)
  • Most of the above mentioned programs are single threaded, thus high clock speed is important.

MEMORY: 32 Gb (or maybe even 64) of memory
  • ECC or non-ECC depending on the CPU.

GRAPHICS: AMD FirePro v5900 2Gb
  • Have no experience here, but read somewhere that an intermediate card would be preferable.

HARD DRIVES: 1x128 Gb HP SSD for software install, and 2x256 HP SSD in RAID1 for working directory.
  • the waiting time is often dependent on the loading of cases 1-4 Gb's in size. RAID1 to have some sort of security.

I really need some inputs here, all suggestions and your personal experiences are more that welcome. I am quite new in the industry (finished my masters this summer), so all constructive comments are welcome. I am especially interested in the choosing of the hard drives: Are SSDs the way to go or is SAS 10k or 15k drivers better, would another raid solution be better bang-for-the-buck (RAID5/10) ?

Thank you very much in advance. Feel free to link to other posts/forum discussion which might have interest.

Last edited by larsenmm; January 5, 2013 at 09:43.
larsenmm is offline   Reply With Quote

Old   January 5, 2013, 19:37
Default
  #2
Senior Member
 
Join Date: Oct 2009
Location: Germany
Posts: 636
Rep Power: 22
abdul099 is on a distinguished road
HDD:
I think most of the loading times is not the "reading from HDD" but more processing of the loaded data in the current session.

I'm having a 10k rpm SAS HDD in my workstation and never had HDD related issues with loading times (in STAR-CCM+). This HDD is fast enough to read even big files with let's say up to 15Gb in 1.5 minutes, so I'm not sure if it's worth to spend the money for a SSD - especially when you're running a RAID 1.
Just a side note, even the HDD in my private PC, which is a 7200 rpm SATA WD black series, needs less than 3 minutes to load a file of up to 4GB. It reads more than 90 MB/s which is ~5.4GB / min. The performance NEVER dropped so low that it would take just 3 minutes to load a 4GB file.

Please note, a RAID 1 will increase only read performance. Write performance will be as good as a single disk. When there's another backup solution available at your company, I would not go for a RAID 1. Just buy a single disk then, that should be enough. I had no sudden disk failure so far since I've used my first PC 20 years ago.


CPU:
Right. Single core performance will be the most important point. Nothing to add here.

Memory:
When money is an issue, get non-ECC RAM since it's cheaper. Maybe you can get a faster CPU instead...

Graphics:
I don't have any experience with Ansys products and graphics performance is pretty much related to the software. So I'm sorry, I can't give any advice here.


And please don't forget, it's CFD. You're handling more or less big models. That takes some time, and sometimes you have to wait. Even with a fast CPU, you won't be able to shortcut the meshing time from 60 min to 10 min. So be patient and try to use your time as efficient as possible - there will be no miracle even with a new machine...
larsenmm and rgd like this.
__________________
We do three types of jobs here:
GOOD, FAST AND CHEAP
You may choose any two!
abdul099 is offline   Reply With Quote

Old   January 6, 2013, 08:53
Wink
  #3
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Quote:
Originally Posted by abdul099 View Post
HDD:
I think most of the loading times is not the "reading from HDD" but more processing of the loaded data in the current session.

I'm having a 10k rpm SAS HDD in my workstation and never had HDD related issues with loading times (in STAR-CCM+). This HDD is fast enough to read even big files with let's say up to 15Gb in 1.5 minutes, so I'm not sure if it's worth to spend the money for a SSD - especially when you're running a RAID 1.
Just a side note, even the HDD in my private PC, which is a 7200 rpm SATA WD black series, needs less than 3 minutes to load a file of up to 4GB. It reads more than 90 MB/s which is ~5.4GB / min. The performance NEVER dropped so low that it would take just 3 minutes to load a 4GB file.

Please note, a RAID 1 will increase only read performance. Write performance will be as good as a single disk. When there's another backup solution available at your company, I would not go for a RAID 1. Just buy a single disk then, that should be enough. I had no sudden disk failure so far since I've used my first PC 20 years ago.
  • I hear what you say, and I agree that it is not only the HDD that causes the long load times. Another reason for me wanting the SSD (I did not mention it above) for at least the system drive, is the smoothness you get from running an SSD. I open and close programs quite often, and enjoy working on my present laptop with SSD compared to the older Z600 workstation with an ordinary HDD. I have no experience working with fast SAS disks, and if you be able to get the same speedy feeling from a SAS disk ?

    I came to think about the system disk last night, and I think the size of the system drive is to low. With 32-64 gb of ram i loose more or less the same amount of GB to the Windows page file (as far as i can figure out). On top of that it seems to be a good idea to keep around 20 percent of your SSD's space free (depending on the quality of the garbage collection of the SSD) which very much limits the space available on a 128 GB disk. With windows 7 and various programs i can easily get into problems. Thus a 180 or 256 GB SSD would be better I think. Am I correct in my thinking ?

Quote:
Originally Posted by abdul099 View Post
CPU:
Right. Single core performance will be the most important point. Nothing to add here.
  • Ok, that was my experience also.

Quote:
Originally Posted by abdul099 View Post
Memory:
When money is an issue, get non-ECC RAM since it's cheaper. Maybe you can get a faster CPU instead...
  • As far as i can figure out browsing different forums, ECC memory is important on our new cluster which will do all the number crunching. On top of that it has a larger number of hardware parts (memory modules, CPU's) that have to work flawlessly. In my case for the workstation, which is not meant for heavy number crunching,
    I do not think it is worth spending the extra money on ECC. Except in the case that i want the CPU with the highest clock, since that is a Xeon E3-1290v2 which require ECC memory, right ?

Quote:
Originally Posted by abdul099 View Post
Graphics:
I don't have any experience with Ansys products and graphics performance is pretty much related to the software. So I'm sorry, I can't give any advice here.
  • ok, fair enough

Quote:
Originally Posted by abdul099 View Post
And please don't forget, it's CFD. You're handling more or less big models. That takes some time, and sometimes you have to wait. Even with a fast CPU, you won't be able to shortcut the meshing time from 60 min to 10 min. So be patient and try to use your time as efficient as possible - there will be no miracle even with a new machine...
  • I understand that I will not avoid the waiting . But even a smaller speedup would be welcome. The last case I was working on, was with a very bad quality CAD (from solid works) which induced a large amount of errors in ANSYS Meshing. Thus I had to reiterate my CAD cleaning/Meshing numerous times. This process ended up being to slow and I had to change tactics and go with a much coarse mesh due to time limits. Doing Inflation layers (wall refinements) in ANSYS Meshing, require very good quality CAD. And even then, it is a very slow process.

    I am not expecting miracles, but I do expect some sort of speedup.
rgd likes this.
larsenmm is offline   Reply With Quote

Old   January 6, 2013, 16:40
Default
  #4
Senior Member
 
scipy's Avatar
 
Alex Pasic
Join Date: Aug 2011
Location: Croatia
Posts: 199
Rep Power: 16
scipy is on a distinguished road
Send a message via Skype™ to scipy
Quote:
Originally Posted by larsenmm View Post
I have no experience working with fast SAS disks, and if you be able to get the same speedy feeling from a SAS disk ?
They're still mechanical disks, regardless of their rotation rate. What you are experiencing with the SSD is the ultra-low latency and response that comes with flash memory. Get the SSD.

Quote:
Originally Posted by larsenmm View Post
I came to think about the system disk last night, and I think the size of the system drive is to low. With 32-64 gb of ram i loose more or less the same amount of GB to the Windows page file (as far as i can figure out). On top of that it seems to be a good idea to keep around 20 percent of your SSD's space free (depending on the quality of the garbage collection of the SSD) which very much limits the space available on a 128 GB disk. With windows 7 and various programs i can easily get into problems. Thus a 180 or 256 GB SSD would be better I think. Am I correct in my thinking ?

Since you're not footing the bill, just get the biggest one you can. However, the reasons you list are invalid. The paging file is only "needed" if you plan on experiencing very frequent system crashes as the memory contents gets dumped there. The first thing I do after installing Windows is to disable the hybernation (and that deletes the huge hybernation file from C:\ root) and decrease the page file size to 64-128 mb. Otherwise nearly 40 gb are gone from the SSD right after installing Windows (and I have 32 gb of RAM). My OS disk is a 128 gb OCZ Vertex4 and even with EVERYTHING installed (Windows, Office, Workbench 14.5, Corel Technical, full Acrobat and Photoshop, Mathcad, Tecplot, SolidWorks, Rhinoceros and other small miscellaneous programs) there is still 66 gb of free space (119 gb recognized size of the SSD), so it's about 45 % full. Plenty of space left to keep the "work in progress" working directories on the SSD, too. I move them to a 7200 rpm disk once everything is done.


Quote:
Originally Posted by larsenmm View Post
Except in the case that i want the CPU with the highest clock, since that is a Xeon E3-1290v2 which require ECC memory, right ?
No, that is not right. Xeon *can* work with ECC memory, while other Desktop cpus cannot, but it doesn't require it to work. In fact, even your cluster probably doesn't require it. ECC is used in "mission critical" deployments where number-crunching time is in measured in weeks/months for a single uninterrupted calculation (no reboots or release of memory), because in those situations even a single error bit in the memory can multiply and propagate until it crashes the whole process. If your calculations are up to a few days in length of uninterrupted runtime, then ECC is not really needed.


Quote:
Originally Posted by larsenmm View Post
I understand that I will not avoid the waiting . But even a smaller speedup would be welcome. The last case I was working on, was with a very bad quality CAD (from solid works) which induced a large amount of errors in ANSYS Meshing. Thus I had to reiterate my CAD cleaning/Meshing numerous times. This process ended up being to slow and I had to change tactics and go with a much coarse mesh due to time limits. Doing Inflation layers (wall refinements) in ANSYS Meshing, require very good quality CAD. And even then, it is a very slow process.
This is where you should change your ways and not rely on certain types of software as much. For "dirty" geometry, even though I don't particularly consider anything that comes out of SolidWorks very dirty (since it's usually NURBS) - but I haven't seen your models so, I'll trust you; you should use applications like ICEM CFD and/or TGrid. ICEM probably has more potential for geometry cleanup and repair, while if your geometry is just HOPELESS, then TGrid has very good wrapping capabilities (if you don't mind the loss of some very small details). Combining those two, for example - using ICEM CFD to produce a quality surface mesh and then using TGrid (now called FLUENT Meshing in V14.5) for prisms and a quality volume mesh is probably unbeatable. You could also tetra mesh your regions of "complexity" in ICEM CFD and pretty easily do the rest of the "simple, rectangular" geometry with ICEM Hexa and then merge the two regions together, export mesh to CFX and boom. Much less waiting and going back and forth between CAD fixing and ANSYS Meshing.

P.S. Nice to see someone who understands quoting and uses it. It's a nice change from the norm. :P
larsenmm likes this.
scipy is offline   Reply With Quote

Old   January 8, 2013, 02:09
Default
  #5
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
scipy: Thank you for your reply, very informative.

Your points about the SSD and memory sounds as they make good sense, so I will try to follow those. Do you (or anyone else) have any inputs regarding SSD's and raid (0, 1 or 10).
  • Would the TRIM function work in raid ?.
  • From experience I think I would like 100-200 Gb as a minimum for a working directory. Is there a particular reason why I should or should not split windows drive and working directory up on two physical hard drives?
  • Anything I have not thought of?


Quote:
This is where you should change your ways and not rely on certain types of software as much. For "dirty" geometry, even though I don't particularly consider anything that comes out of SolidWorks very dirty (since it's usually NURBS) - but I haven't seen your models so, I'll trust you; you should use applications like ICEM CFD and/or TGrid. ICEM probably has more potential for geometry cleanup and repair, while if your geometry is just HOPELESS, then TGrid has very good wrapping capabilities (if you don't mind the loss of some very small details). Combining those two, for example - using ICEM CFD to produce a quality surface mesh and then using TGrid (now called FLUENT Meshing in V14.5) for prisms and a quality volume mesh is probably unbeatable. You could also tetra mesh your regions of "complexity" in ICEM CFD and pretty easily do the rest of the "simple, rectangular" geometry with ICEM Hexa and then merge the two regions together, export mesh to CFX and boom. Much less waiting and going back and forth between CAD fixing and ANSYS Meshing.
I would not mind to change my ways. Your way sounds interesting and I will try to see if I am limited through our license (ICEM + TGrid) and if I can try "your ways" . I will have a meeting/telecon with our software supplier in the nearest future to discuss this problem in particular, as they have suggested to look through my ways and our geometries to see if it can be done in an easier way.

Quote:
P.S. Nice to see someone who understands quoting and uses it. It's a nice change from the norm. :P
Thank you

And thanks all in advance for your inputs

Last edited by larsenmm; January 22, 2013 at 02:08. Reason: misspellings
larsenmm is offline   Reply With Quote

Old   January 8, 2013, 03:47
Default
  #6
Senior Member
 
scipy's Avatar
 
Alex Pasic
Join Date: Aug 2011
Location: Croatia
Posts: 199
Rep Power: 16
scipy is on a distinguished road
Send a message via Skype™ to scipy
Quote:
Originally Posted by larsenmm View Post
Would the TRIM function work in raid ?
TRIM will not work in RAID for SSDs.

Quote:
Originally Posted by larsenmm View Post
Is there a particular reason why I should or should not split windows drive and working directory up on two physical hard drives?
No reason. I keep the working dirs on the SSD (C:\) purely because of faster read/writes. If I had another SSD I could keep them there with no penalties.

In any case, most of the ANSYS software compresses the case and data files within Workbench by default (via gzip or 7zip), so the bottleneck for reads/writes is not the hard drive anyway (it hasn't been even on SATA 2 SSDs).
rgd likes this.
scipy is offline   Reply With Quote

Old   January 16, 2013, 02:20
Default
  #7
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Quote:
TRIM will not work in RAID for SSDs.
I though i saw some news on the subject about TRIM and SSD's in Raid so I hoped someone in here had experience with it. I will see what else i can find on the subject.

Instead of two SSD's in raid what would be alternative ? Could one install a regular HDD or SSD (on non-raid config) which then mirrors the primary SSD once in a while. Like once a day or maybe twice depending how much load that would be put on the workstation and if it would slow down my work. This would be setup using some kind of software is my guess. Anyone ?

Quote:
No reason. I keep the working dirs on the SSD (C:\) purely because of faster read/writes. If I had another SSD I could keep them there with no penalties.

In any case, most of the ANSYS software compresses the case and data files within Workbench by default (via gzip or 7zip), so the bottleneck for reads/writes is not the hard drive anyway (it hasn't been even on SATA 2 SSDs).
I do use the "Archive" option in Workbench but not to a great extend. I guess as with regular compression jobs, the bottleneck is the CPU.
larsenmm is offline   Reply With Quote

Old   January 16, 2013, 09:14
Default
  #8
Senior Member
 
Robert
Join Date: Jun 2010
Posts: 117
Rep Power: 17
RobertB is on a distinguished road
One interesting thing about SSDs is that they have a life (as opposed to a MTBF). As the devices are becoming denser/smaller sized it is dropping. The article below discusses the issue for an Intel drive.

http://www.anandtech.com/show/6462/i...on-the-ssd-335

While the life is very high for normal usage, the big files and potentially frequent writes in analysis usage may cause them to wear out relatively rapidly.
RobertB is offline   Reply With Quote

Old   January 16, 2013, 15:15
Default
  #9
Senior Member
 
scipy's Avatar
 
Alex Pasic
Join Date: Aug 2011
Location: Croatia
Posts: 199
Rep Power: 16
scipy is on a distinguished road
Send a message via Skype™ to scipy
These concerns were mostly aimed at SSD usage in data centers with REALLY high frequency of writes to the drive. The data industry as such pressured the SSD makers to make drives that would survive for 5 years in such operating conditions - and this is why I recommended a synchronous (not asynchronous) MLC NAND SSDs. The mentioned Vertex 4 (and/or Corsair's Force GT and Intel's 520 series of drives) comes with 5 years warranty just to soothe concerns from users like yourself, who were worried about life expectancy of a SSD.

So, with CFD usage, even if it was fully transient running 100 % of the time, saving .cas and .dat files every few iterations, these SSDs should not lose any noticable amount of data storage/space on them nor performance.

In other words, don't worry. In less than 5 years these are going to be as outdate as SATA-150 drives are today and we will all move to SATA5 and unbelivable transfer speeds, before your current SSD dies of repeated read-writes.
scipy is offline   Reply With Quote

Old   January 16, 2013, 17:38
Default
  #10
New Member
 
Frederik
Join Date: Apr 2010
Posts: 18
Rep Power: 16
fchan is on a distinguished road
Quote:
Originally Posted by larsenmm View Post
I though i saw some news on the subject about TRIM and SSD's in Raid so I hoped someone in here had experience with it. I will see what else i can find on the subject.
SSDs can be trimed in RAID 0 if you use gen-7 or newer Intel chipset. More details here: http://www.anandtech.com/show/6161/i...rds-we-test-it
fchan is offline   Reply With Quote

Old   January 18, 2013, 17:53
Default
  #11
Senior Member
 
Join Date: Oct 2009
Location: Germany
Posts: 636
Rep Power: 22
abdul099 is on a distinguished road
Quote:
Originally Posted by scipy View Post
These concerns were mostly aimed at SSD usage in data centers with REALLY high frequency of writes to the drive. The data industry as such pressured the SSD makers to make drives that would survive for 5 years in such operating conditions - and this is why I recommended a synchronous (not asynchronous) MLC NAND SSDs. The mentioned Vertex 4 (and/or Corsair's Force GT and Intel's 520 series of drives) comes with 5 years warranty just to soothe concerns from users like yourself, who were worried about life expectancy of a SSD.

So, with CFD usage, even if it was fully transient running 100 % of the time, saving .cas and .dat files every few iterations, these SSDs should not lose any noticable amount of data storage/space on them nor performance.

In other words, don't worry. In less than 5 years these are going to be as outdate as SATA-150 drives are today and we will all move to SATA5 and unbelivable transfer speeds, before your current SSD dies of repeated read-writes.
Right, I think SSD are more reliable today as they're called to be. But on the other hand, you can have carbon brake disks on a Ford T-model - but does it help in making it faster?
For most simulations run out there in the wilderness of the CFD-world, disk write speed is not the issue slowing down the simulations. When I run a case on my desktop at home and I save a 1Gb sim-file file every 500 iterations, it would take less than 10 seconds with my HDD and maybe 4 seconds with a SSD. This would occur once every 8 hours. So I save a horrible amount of time, right? Since it would take only 7:59:54 instead of taking 8:00:00.
Sure, it would become more important when the frequency of writing data to the hard disk increases, but I'm a consultant running CFD simulations since several years, for all kind of industries - and disk speed was NEVER an issue although I've never in my life used a SSD.

So as a conclusion, I would say: SSD's make sense for the system disk, since program startup will be accelerated a lot (less than 10 seconds, since this is how much it takes with my HDD). Once a day, before I start the simulation. And people having too much money also might use them for the main data storage (although they could also transfer the excess money to me, I'll appreciate it ). But having a SAS HDD, or for a home workstation even a fast SATA HDD, should be more than sufficient. Up to you what you choose, I'll spend the money to get more beer for applied fluid dynamics
larsenmm likes this.
__________________
We do three types of jobs here:
GOOD, FAST AND CHEAP
You may choose any two!
abdul099 is offline   Reply With Quote

Old   January 22, 2013, 02:14
Default
  #12
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Quote:
Originally Posted by scipy View Post
TRIM will not work in RAID for SSDs.

No reason. I keep the working dirs on the SSD (C:\) purely because of faster read/writes. If I had another SSD I could keep them there with no penalties.
I was thinking more in the line of downtime, reinstallation problems, data security and ease of use. Having windows and programs on a separate disk means I can have a major software fuckup (blue screen of death etc), reinstall (hopefully using an image) and be back online in no time... Or does this happen so seldom so i will not notice the advantages ?

Kind regards
Martin
larsenmm is offline   Reply With Quote

Old   February 6, 2013, 10:36
Default
  #13
Member
 
Kim Bindesbøll Andersen
Join Date: Oct 2010
Location: Aalborg, Denmark
Posts: 39
Rep Power: 16
bindesboll is on a distinguished road
My system (used for SpaceClaim, ANSYS CFX Pre, Post and Solver):
Dell T5500, Xeon 5687 3.6 GHz (quad-core), 24 GB 1333 MHz non-ECE Ram, 2 x 250 GB SSD, ATI FirePro V7900.

Meshing:
I have tried nVidia QuadroPro 2000 graphics card with meshing and wanted something faster, as rotating and similar could be with several seconds of delay, when showing a very fine mesh. However, I think that my present FirePro V7900 is not that much better in this regard, even though it should be much faster. So I think that graphic issues might depend more on the processor and RAM speed as the graphics card does not seem to do it all.
Meshing of 6 mio. cells takes 3-5 min.
Meshing is not that good to utilise multiple cores. Only two cores show more than 50% load when meshing. So I think that bandwith between CPU and RAM is limiting.

CFX Pre:
Loading a single phase, multi species, 11 mio. cells case takes 5-10 min. CPU loading is low during loading, so I would guess that it is SSD bandwidth limited.

My recommendations:
Go for high single core performance as multiple cores are not utilised efficiently.
Go for high memory bandwidth.
Go for SSD.
Dont dont spend all the money on graphics card.

Best regards
Kim
larsenmm likes this.
bindesboll is offline   Reply With Quote

Old   February 6, 2013, 12:30
Default
  #14
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Quote:
Originally Posted by bindesboll View Post
My system (used for SpaceClaim, ANSYS CFX Pre, Post and Solver):
Dell T5500, Xeon 5687 3.6 GHz (quad-core), 24 GB 1333 MHz non-ECE Ram, 2 x 250 GB SSD, ATI FirePro V7900.
Hi Kim, thank you very much with some spot-on info. If I may ask, how do you run the 2 x SSDs, in a raid or just normal setup ?


Quote:
Meshing:
I have tried nVidia QuadroPro 2000 graphics card with meshing and wanted something faster, as rotating and similar could be with several seconds of delay, when showing a very fine mesh. However, I think that my present FirePro V7900 is not that much better in this regard, even though it should be much faster. So I think that graphic issues might depend more on the processor and RAM speed as the graphics card does not seem to do it all.
Meshing of 6 mio. cells takes 3-5 min.
Meshing is not that good to utilise multiple cores. Only two cores show more than 50% load when meshing. So I think that bandwith between CPU and RAM is limiting.
I am sorry to hear that even with your setup you have problems with speed in Ansys Meshing (I assume that is the one you use). The setup which I was aiming for was very similar. I think then I will try to contact ANSYS or out supplier themselves to get feedback on what they would recommend. It might not be within my budget, but at least now I know that the planned V5900 does not make everything nice and smooth.

Regarding cells, we are talking number of cells not nodes, correct ? That would make a huge difference, and I was told to use the number of nodes if I was comparing with Fluent or AVL Fire as there is a difference in the solvers. As far as I remember, CFX is a node based solver, where Fluent and AVL Fire are cell centered.

Quote:
CFX Pre:
Loading a single phase, multi species, 11 mio. cells case takes 5-10 min. CPU loading is low during loading, so I would guess that it is SSD bandwidth limited.
Okay, very interesting to know. If I get the possibility, I will try with both single SSD, Raid 1 and Raid 0 to see how important the SSD bandwidth is.

Quote:
My recommendations:
Go for high single core performance as multiple cores are not utilised efficiently.
Go for high memory bandwidth.
Go for SSD.
Dont dont spend all the money on graphics card.

Best regards
Kim
Below I have listed the hardware which I am aiming for. I think that fits more or less with your recommendations, and what else people in here have been saying.

The workstation is a HP Z220.
- HP Core i7-3770 3.4GHz 4C (highest clocked i7 available, Turboboost to 3.9Ghz)
- HP 32GB DDR3-1600 (4x8GB) nECC (maximum amount possible, 2 memory channels, 2 sockets per channel. Is this a major issue, if I am not running simulations myself ?)
- HP 256GB SATA 1st SSD (main OS disk, programs, user files)
- HP 256GB SATA 2nd SSD (Raid 1, disk 1, project files for running project, Raid 1 and 0 officially supported by HP)
- HP 256GB SATA 3rd SSD (Raid 1, disk 2, project files for running project, Raid 1 and 0 officially supported by HP)
- HP AMD FirePro V5900 2GB Graphics
- HP Z220 CMT 400W 90% Efficient
- HP 3y 4h 13x5 Onsite Only HW (so i can blame HP if something does not work)

Some of my colleges have bought their own consumer hardware, and assembled a workstation from that. This can be done for maybe half the price as a HP Workstation, but with the amount of errors/problems seen on consumer SSDs, too hot CPU's and the like, getting the responsibility off my back is relatively important. I do not mind working with hardware, but If I am on a strict deadline at work, I would rather do my work than trying to fix my computer!

Please again feel free to comment, the workstation have not been purchased yet but hopefully will be within a short time.

Thanks in advance

Best regards
Martin

Last edited by larsenmm; February 6, 2013 at 12:49. Reason: Realised that the Z220 only support dual channel memory and not 4 channels as the Z420!
larsenmm is offline   Reply With Quote

Old   February 12, 2013, 09:46
Default
  #15
Member
 
Kim Bindesbøll Andersen
Join Date: Oct 2010
Location: Aalborg, Denmark
Posts: 39
Rep Power: 16
bindesboll is on a distinguished road
Quote:
Originally Posted by larsenmm View Post
If I may ask, how do you run the 2 x SSDs, in a raid or just normal setup ?
Just normal setup? Now and then I archieve my work on a network drive.

Quote:
Originally Posted by larsenmm View Post
I am sorry to hear that even with your setup you have problems with speed in Ansys Meshing (I assume that is the one you use). The setup which I was aiming for was very similar. I think then I will try to contact ANSYS or out supplier themselves to get feedback on what they would recommend. It might not be within my budget, but at least now I know that the planned V5900 does not make everything nice and smooth.
Yes, I use ANSYS Meshing.
My point is that what-ever graphics card you choose, Meshing graphics can still be heavy, as (apparently) it does not all depend on the graphics card. However, I dont experience it as a big problem. Only when showing large dimensions of a fine mesh, and rotating or similar. Nothing that obstructs my work at all.

Quote:
Originally Posted by larsenmm View Post
Regarding cells, we are talking number of cells not nodes, correct ? That would make a huge difference, and I was told to use the number of nodes if I was comparing with Fluent or AVL Fire as there is a difference in the solvers. As far as I remember, CFX is a node based solver, where Fluent and AVL Fire are cell centered.
You are right. I stated in cells as everybody else does.

Quote:
Originally Posted by larsenmm View Post
The workstation is a HP Z220.
- HP Core i7-3770 3.4GHz 4C (highest clocked i7 available, Turboboost to 3.9Ghz)
- HP 32GB DDR3-1600 (4x8GB) nECC (maximum amount possible, 2 memory channels, 2 sockets per channel. Is this a major issue, if I am not running simulations myself ?)
As far as I know the Core i-7 only has 2 memory channels, whereas Xenon E5 has 4 memory channels. As memory bandwidth is critical you should go for a Xenon and utilize all memory channels (8x4GB instead of 4x8GB).


Best regards
Kim Bindesbøll
larsenmm likes this.
bindesboll is offline   Reply With Quote

Old   February 12, 2013, 10:49
Default
  #16
Senior Member
 
Joern Beilke
Join Date: Mar 2009
Location: Dresden
Posts: 540
Rep Power: 20
JBeilke is on a distinguished road
Memory bandwidth becomes important when you use more than one core. For pre and post it should not matter.

The i7 can use faster memory than the Xeons.
larsenmm likes this.
JBeilke is offline   Reply With Quote

Old   February 12, 2013, 11:28
Default
  #17
New Member
 
Martin Larsen
Join Date: Nov 2012
Location: Middelfart, Denmark
Posts: 9
Rep Power: 14
larsenmm is on a distinguished road
Thanks for your replies bindesboll and JBeilke

Quote:
Originally Posted by JBeilke View Post
Memory bandwidth becomes important when you use more than one core. For pre and post it should not matter.

The i7 can use faster memory than the Xeons.
This was more or less my understanding also.

bindesboll, you say critical but critical for what work exactly? I agree to your statement when we are talking about running the actual solving on the specific workstation. But is it also critical for Post/Pre work ?

The reason why I am so interested in this, is that i need to go from a HP Z220 to a Z420 to get 4 channel memory, and i need to get an idea of the performance gain compared to the extra cost...
larsenmm is offline   Reply With Quote

Old   February 13, 2013, 05:16
Default
  #18
Member
 
Kim Bindesbøll Andersen
Join Date: Oct 2010
Location: Aalborg, Denmark
Posts: 39
Rep Power: 16
bindesboll is on a distinguished road
I was wrong regarding Pre and Post being memory bandwidth critical. I agree with JBeilke, that most Pre and Post operations are not multi-core heavy to that high degree. I guess operations are typically disc speed or single-core-performance critical instead.
larsenmm likes this.
bindesboll is offline   Reply With Quote

Reply

Tags
hardware, post procesing, pre-processing, ssd


Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are Off
Pingbacks are On
Refbacks are On


Similar Threads
Thread Thread Starter Forum Replies Last Post
different CFX Pre and Post mesh region mactech001 CFX 9 April 11, 2010 22:08
Objects in PRE and POST ? Felix CFX 0 January 16, 2008 16:53
clarification in LS Pre post David Main CFD Forum 0 July 19, 2007 14:02
Instable Post, Pre and Build Pascale Fonteijn CFX 5 December 23, 2003 07:31
PRE and POST processing package Astrid Barros Main CFD Forum 6 November 16, 2000 11:31


All times are GMT -4. The time now is 22:44.