Low-end Scientific Workstation

[Alumbros]

1. I am a student of mathematics, specifically combinatorics. Recently, I have found it worthwhile perform computations and enumerate combinatorial objects. I mostly compute with integers rather than floating point. Some of the computations run for weeks at a time. For example, I often need to find independent sets in a graph.

I am currently using an Intel Core 2 Duo E7400 with 4GB of ram, and I feel an upgrade is in order.

Therefore, I would like to build a single-cpu workstation with several cores and at least 16GB of ECC memory. A single hard drive will suffice. I would also like a GPU for GPU-accelerated computations, but this is not a priority. If possible, the GPU should also have ECC memory; however, if this proves too expensive, I can live without it.

I certainly want a workstation case rather than a server case. I will probably also end up using this system for general computing tasks, such as web browsing, etc. Note that I am doing this on my own, without any support from an IT department.

2. Budget: $1200 - $1500
3. USA
5. I would prefer an Intel Xeon processor. However, any cpu/mobo combination that supports ECC ram is acceptable.
6. I can reuse a monitor, keyboard and mouse.
7. Not overclocking, stability and reliability are more important
8. 1920x1080 at least
9. I would like to start using this system in 4 to 6 weeks. However, I can wait longer.
X. I do not need to purchase any software. I will run Linux as an OS. I use Sage, GAP, and other mathematical software as necessary.

 

[college_student]

Hi Alumbros,

Why are you interested specifically in a Xeon processor? I recently built a cluster for scientific computing and a 3770K (even before it was overclocked modestly to 4.2 GHz) seemed to outperform some of the latest Xeon CPU clusters running the same signal processing routines.

Also, in my somewhat limited experience, the ECC memory doesn't seem to matter all that much. But please don't quote me on this. I defer this to others who are more knowledgeable. But in my experience it really doesn't matter all that much.

Basically it would be tough to build a really good scientific computing workstation if you're spending a lot of money on stuff like Xeon processors and ECC memory. But you can build a really strong one if you go with consumer grade stuff like i7, etc..

Thanks,
-DV

 

[Sleepingforest]

Maybe the lowered error rate is worthwhile to him? Since the calculations run over weeks, power consumption and error rate matter quite a lot more than a mere hour long calculation.

Do you know what your programs are running on (such as OpenCL, OpenGL, CUDA, etc)? Because that will make a significant difference in the GPU you need.

As a preliminary recommendation:
CPU: Intel Xeon E3-1230 V2 3.3GHz Quad-Core Processor ($239.99 @ Newegg)
Motherboard: Supermicro Xeon E3 Motherboard ($193.98 @ Newegg)
RAM: Crucial 2x8GB ECC Registered RAM ($133.99 @ Newegg)
Storage: Seagate Barracuda 2TB 3.5" 7200RPM Internal Hard Drive ($99.99 @ Amazon)
Video Card: ATI FirePro V5900 2GB Video Card ($430.98 @ Newegg)
Case: Corsair 200R ATX Mid Tower Case ($59.99 @ Amazon)
Power Supply: Corsair Builder 430W 80 PLUS Bronze Certified ATX12V Power Supply ($39.99 @ Newegg)
Optical Drive: Lite-On iHAS124-04 DVD/CD Writer ($15.99 @ Newegg)
Total: $1214.90 (before $20 in rebates).

 

[Alumbros]

Thanks for the replies. I am a neophyte when it comes to GPU computing. However, CUDA seems to be popular for the kinds of problems I am interested in.

I would sleep better at night knowing ECC was on, but I would consider dropping it.

 

[Sleepingforest]

My build is ECC all the way through (RAM, VRAM, motherboard supported, plus a Xeon). However, it has an AMD FirePro rather than an Nvidia Quadro (Nvidia has proprietary control over CUDA).

You can get an Nvidia Quadro 2000 with 1GB GDDR5 for $400 from Newegg.

 

[college_student]

Thanks for the replies. I am a neophyte when it comes to GPU computing. However, CUDA seems to be popular for the kinds of problems I am interested in.

I would sleep better at night knowing ECC was on, but I would consider dropping it.

I really would urge you to consider going for the consumer grade parts. I went through the same process you are when I built the cluster. When you look at enough benchmarks you realize that you are getting a significantly better bang for the buck compared to the server grade parts. If you're not thinking about buying a dual-CPU mobo, or have no interest in that, then you really can't go wrong with the 3770K. Also, the FirePro etc, in my opinion, are a lot more expensive for the same performance of a consumer card. I've already told you my thoughts on ECC (I know plenty of people who work in Academia and Industry who do just fine with non-ECC memory). But it's your choice, and I fully support piece of mind!

Just my thoughts,
-DV

 

[Sleepingforest]

I'm fairly certain that a memory error would be crippling for a long term project or calculation, and with the number of calculations that go on during a week(s) long number crunch, I'd rather be safe than sorry--thus ECC RAM.

In this case, since it will be on constantly and presumably under a fair amount of load, overclocking and the implied energy requirement is not ideal (even if you don't pay for the electricity). Furthermore, the i7 actually costs more than the Xeon I recommended. Finally, I don't know how well Alumbros's programs will scale better with core/thread count or clockspeed, but if it is the former, then overclocking would not be very helpful anyway.

Ultimately, it is up to Alumbros to decide. You (college_student) do bring up some good price/performance points, but if the safety fits in the budget, why not?

 

[Cerb]

Check out the Intel S1200BTSR, and ASUS P8B-X, for mobos other than Supermicro's. Also, consider moving up to the E3-1240 V2. It's only 100MHz, but your budget can take it with no problem.

In general, Sleepingforest's build is right on. The cost to get LGA2011 is high. The cost to get registered ECC is high (not the RAM, but the CPUs and mobos). The cost to go to unbuffered ECC (no address parity) is not high, and fits right in your budget.

I need sleep, so I'll leave you to find and read it, but go check out Google's RAM study. In particular, note that DIMM/chip count, and memory utilization (reads and writes) correlated positively with errors, and that most computers that had 1 error, would have more. Then, consider that, without ECC, you have no way to even tell if an error happened, without performing the same operations multiple times, and comparing the results (awfully inefficient, compared to getting ECC RAM). Chances are, it hasn't, but your budget is quite reasonable for it, as is your use case, so why not do it and have a little more peace of mind?

Another option, of course, would be Dell.
http://outlet.us.dell.com/ARBOnline...arch.aspx?c=us&l=en&s=dfb&cs=28&puid=87597d31
There are several around that price range, with the same CPU, W7 Pro, and 16GB registered ECC. Not bad at all, really, though the GPU wouldn't be much good for computation. The ones with Quadro 2000 cards or better seem to go over-budget, or have slower CPUs.

 

[mfenn]

<--------- HPC admin here.

Once your simulations reach a certain time threshold (~ 48 hours is my metric) you should really start building checkpointing into your code. ECC memory is all well and good, but there are many many other reasons that a simulation can fail (OS bug, cpu overheats, power outage, the knucklehead behind the keyboard accidentally hitting Ctrl-C etc.) beyond memory corruption. To make any sort of reasonable forward progress on a weeks-long job, you must be able to cope with failure and restart.

With regard to bit instability affecting numerics, you will also want to build in some sort of check condition into your code. That way, you your code can detect whether or not it is going off the rails and abort if so. Then you examine your last checkpoints and figure out where it all went wrong.

I'm not saying that you shouldn't get ECC (it's not super expensive), but I am saying that it isn't a panacea. Good, failure-tolerant code is even more important than the world's fanciest memory.

 

[Rikard]

Don't you have access to some computing centre? If you need a week to run your calculations you really should look into distributed computing rather than buying hardware. Factorize the calculations and run them in parallel. (That is what I do for my calculations, for example this week I submitted 13531 jobs since Monday and that is just a normal week.)

 

[TigCobra]

<--------- HPC admin here.

Once your simulations reach a certain time threshold (~ 48 hours is my metric) you should really start building checkpointing into your code. ECC memory is all well and good, but there are many many other reasons that a simulation can fail (OS bug, cpu overheats, power outage, the knucklehead behind the keyboard accidentally hitting Ctrl-C etc.) beyond memory corruption. To make any sort of reasonable forward progress on a weeks-long job, you must be able to cope with failure and restart.

With regard to bit instability affecting numerics, you will also want to build in some sort of check condition into your code. That way, you your code can detect whether or not it is going off the rails and abort if so. Then you examine your last checkpoints and figure out where it all went wrong.

I'm not saying that you shouldn't get ECC (it's not super expensive), but I am saying that it isn't a panacea. Good, failure-tolerant code is even more important than the world's fanciest memory.

I agree with mfenn, ECC ram is not important with what you are talking about.