Workstation build suggestion for High Performance Computing (HPC)

I am a researcher and do a lot of HPC stuff. I need to buy a new workstation, so need suggestions from fellow OZ members. Remember, price is of concern here. You can get an idea about how much I can spend on the new workstation by seeing my current workstation configuration. Note that, I intend to use the current and new one in distributed computing (MPI).

My current workstation config is -

  • CPU - Intel Core i7 3770 @ 3.40GHz; 4 core

  • RAM - 16.0GB Dual-Channel DDR3 @ 798MHz

  • Motherboard - Hewlett-Packard 1790

  • Graphics - 1TB NVIDIA Quadro K600

  • Storage - 1.5TB Seagate ST2000DM001-1CH164 ATA Device (SATA)

I'm considering to upgrade the current configuration by adding a SSD -

Now, Should I get a similar Core i7 workstation with higher processor (Intel i7-4930K @ 3.40GHz; 6 core), RAM (32GB) and SSD?
Or should I opt for a Xeon workstation?
Or should I go for a mini cluster solution?

My knowledge is limited regarding HPC hardware, so kindly explain your suggestion also. And No, I can not take advantage of GPU computing because the software I'm using does not support it.

Comments

  • Probably should be asking this question on a more appropriate forum..
    Like OCAU maybe…
    Or ideally a HPC forum.

    • Alternatively there's always whirlpool

  • +3

    for a sec there I thought a 1TB video card might actually exist

    • Phew! Graphics card is 1GB ;)

  • +1

    What is your field of research? What is the software you use? Without knowing more info I don't think our recommendations would be optimally tailored for your workflow.

    • I mostly do Finite Element Analyais. I use Ansys and LS-Dyna. While Ansys is GPU capable, LS-Dyna is not.

      • +2

        Out of ANsys and LS-Dyna, which one is slowing you down the most?

        And once you have that answer, check your PC's resource usage? What is getting maxed out first? CPU, RAM, GPU(if Ansys) or disk access? You want to spend money on the bottleneck first.

        Are your datasets particularly big? Do you really need a 500GB SSD?

        Also, have you done this: http://www.edr.no/index.php/blogg/ansys_bloggen/ansys_tutori…

        • +1

          LS-Dyna is slowing me down. My CPU is maxed out always, 100% (HT off).

          Data set is not really that big, but I intend to use the SSD as Windows drive, so 100GB for Windows + programs and 400GB for simulation files. well, I think 250GB would do it too.

          I will use MPI by joining the current and the new workstation together, total 10 solid cores (No HT).

          1. Does LS-Dyna basically scale linearly as you add new workstations to the cluster?
          2. Can you exclude the cost of licensing from the equation?
          3. Can you confirm that you are really CPU bound?

          If the answer to all those questions is "yes" then perhaps you could consider buying a few of something like this: http://www.graysonline.com/lot/0001-192077/computers-and-it-…

          1. LS-Dyna scales linearly up to 5-6 CPU's, then it depends on the type and size of the problem.

          2. We have site wide license at our campus, so no additional cost for licensing.

          3. Yes, Its CPU bound. Check this screenshot, look at right - http://i.imgur.com/S9FvIdG.jpg

          Few days ago I switched off Hyper Threading (HT) and the performance improved dramatically!

          With 4 3.4GHz core (Turbo Boosting up to 3.9Ghz), I am stuck. How can buying older 2 Core E8400 help this? Also, If I buy these, should not I need to install SSD on them too?

        • With 4 3.4GHz core (Turbo Boosting up to 3.9Ghz), I am stuck. How can buying older 2 Core E8400 help this? Also, If I buy these, should not I need to install SSD on them too?

          The cheap boxes are not to replace your current workstation, but to add more cheap fast CPUs. 2 * 3Ghz worth of processing for less than $150 is pretty cheap.

          If your problem scales linearly by CPU, you can keep adding cheap old boxes.

          If the scaling doesn't extend past 5 or 6 CPUs, then you are better off buying the latest and greatest 6 core CPU you afford.

        • Also, have you explored cloud HPC solutions?

        • Cloud HPC requires payment per hour per core basis. The way academia sees it, you have to have a physical object at your lab (Workstation, Apparatus etc) with the money you have spent on your research.

          Can you suggest some Dual Xeon system within 5000 AUD?

        • Thanks but we can not buy old stuff. It has to be through Uni approved supplier, in our case its HP Australia.

          Can you suggest a dual Xeon configuration for FEA?

        • I suggest you send an RFQ to a HP reseller.

Login or Join to leave a comment