2

Thanks you for this wonderful community.

I have a budget of $10000-15000 for buying a desktop/server which will be running 24*7. I will mainly use it for handling big datasets (20 million observations), running regressions, etc. I mainly use SQL, R, Stata for my data management/ analysis. Some of my codes are quite computationally intensive and might take days to run (think simulations, MCMC, MLE, etc.). I am currently undecided between going the Windows server route or having a linux setup. I am leaning towards Windows right now, due to better software availability. Since this is a significant investment, I would also like this system to be future-proof, i.e., easily upgradable. I also want this system to be able to backup my data on a spare drive.

Request your recommendations. Thanks.

Improve this question
11
  • OK, I think we need a little bit more detail. How much persistent storage do you need? How much RAM do you think will you need? How much redundancy do you want on your data (no RAID, RAID 5?, some more fancy RAID?)? Do you need high-performance networking (eg 10 / 40 GbE?)? How sure do you want to be about up-time (will you need ECC RAM, will you need redundant PSUs, will you need UPS)? How fancy do you want to go on remote-management? Does the noise it generates matter? Will your applications profit from Nvidia GPUs (Quadro / Tesla) or other accelerators like Intel Xeon Phis?
    – SEJPM
    Commented Jun 3, 2017 at 18:13
  • Hello @SEJPM. Thank you for your kind reply. My understanding of hardware is not so great, so I will try to provide the information to the best of my ability, by detailing my end user requirement. I will use this system as my main data storage and data analysis hub. I travel quite a bit (across time zones), so I would be working on this machine remotely. That is why I need the machine to be up 24*7. My analysis usually involves working with big datasets and using models that (sometimes) takes days to converge as they may have to iterate through the entire dataset for each iteration (contd.)
    – gem
    Commented Jun 4, 2017 at 15:15
  • (contd.) hence, a big RAM to store this data in working memory should help with the speed of the analysis? (I am sure that you could see my ignorance by now!). High performance networking should not be an issue, though I want this system to be future-proof. My IT admin tells me that I should get RAID to auto backup my data every day/week. The remote management will involve running codes, data transfer/editing, running multiple codes at the same time in the background. I am not sure about GPUs, sorry. Ideally I would like this system to be low on noise but I could place it in a (contd. 2)
    – gem
    Commented Jun 4, 2017 at 15:23
  • (contd. 2) server room if noise becomes an issue. So, noise should not be a big issue. Storage persistent is not an issue in my case. One more thing that I would like is to have my main data storage drive to be SSD (for speed issues) and the backup drives could be HDD. I have been entertaining the possibility of going with Dell PowerEdge range, but am not sure where to get started. Please let me know if I could provide any more information. Thank you for your help.
    – gem
    Commented Jun 4, 2017 at 15:30
  • @gem How big is an observation? Without context, 20 million sounds like a small number to me. How big is the total amount of data you're working with? Do you need to keep all of it in RAM at the same time, or can individual events be processed independently?
    – CodesInChaos
    Commented Jun 4, 2017 at 17:15

2 Answers 2

Reset to default
1

This question was bumped by the Community user a couple of hours ago. Presumably user "gem" has either purchased a computer or is willing to wait. I'll try to provide a leading edge suggestion that doesn't bust the upper limits of his generous budget. My suggestion only accounts for the cost of a bootable computer and doesn't include the statistical analysis software or the cost of installation and configuration.

I will offer a few suggestions, obviously there are thousands of choices:

Windows or Linux, the hardware is the same - only the software would change.

Statistical Software:

Hardware:

  • Quadro RTX 5000 with 16GB memory - U$2,300

  • Computer: Supermicro A+ SuperServer 4023S-TRT - U$12,377

    • 2x AMD EPYC 7351 Processor 16-core 2.40GHz 64MB Cache (155/170W)

    • 256 GB via 8 DIMMs at 2666 MHz

    • 1.0TB Samsung 960 PRO M.2 PCIe 3.0 x4 NVMe Solid State Drive

    • 10.0TB SATA 6.0GB/s 7200RPM - 3.5" - Seagate Exos X10 Series (Helium)

    • 2x Intel® 10-Gigabit Ethernet Converged Network Adapter X710-T4 (4x RJ45)

    • Microsoft Windows Server 2016 Standard (24-core)

    • 2x Additional 4-core License for Microsoft Windows Server 2016 Standard

    • 5-Device Client Access License (CAL) for Microsoft Windows Server 2016/2012

    • Additional 16-core License for Microsoft Windows Server 2016 Standard

    • 3 Year Depot Warranty (Return for Repair)

Unfortunately that site doesn't have a 'Share this configuration' link but it lets you print out a Spec sheet. If you removed all the Microsoft Windows software and licenses the cost would be U$10,725 without: "Microsoft Windows Server 2016 Standard (24-core)" and "5-Device Client Access License (CAL) for Microsoft Windows Server 2016/2012" and "Additional 2x4-core License for Microsoft Windows Server 2016 Standard".

The Quadro RTX 5000 (16GB with 256-bit memory bus) for U$2,300 isn't available yet and getting a better model involves a steep jump in price with an RTX 6000 (24GB memory and 384 bit bus) for U$6,300 ESP or an RTX 8000 (48GB, also with 384-bit memory bus) for most of your budget at U$10,000 - The JMP Statistical Analysis Software will support those cards and their associated speedup of calculations.

A second 1TB SSD would allow you to RAID the drives for another U$654 but that's punching the roof on the budget. Alternatively two 512GB drives could be had for an extra U$90 over a single 1TB drive. It's almost double the speed versus twice as much on deck.

The hardware is only a suggestion since you'll have to wait until year's end for the GPU, so you might as well wait for the Epyc 8000 series with 7nm technology and likely at least 50% more cores and likely a higher clock rate.

It sort of an in-between time for buying a system for the future. Buying last year's hardware at a considerable savings is another possibility but you lose both the warranty and the big advantages that are promised by year end.

Hardware

Improve this answer
0

Here is a reasonable setup:

If you buy the parts from Amazon, you can assemble for $8600. When we bought from Supermicro reseller (3 years ago), it was $15K.

For building bigger machines, I recommend sites like:

Here you can check what hardware components match each other, customize then buy, or assemble from Amazon (then you have to do the dirty work of assembling and installing OS).

Improve this answer

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.