Making your own server


(Matt) #518

Hi all! As I had promised earlier in this thread, I’ve put together a more detailed blog post on the mATX deep learning rig I spec’d, bought, assembled, configured, and benchmarked. You can find that blog post linked here.

I would really love any feedback you might have on the build and post, as I’m always trying to learn and improve. (and I’d also love to hear any suggestions for how I might continue to put this little guy to work after I run through my current backlog of personal projects!)

Lastly, I just wanted to express my sincere, immense gratitude for all of the posts that come before me in this thread – you’ve been incredibly helpful in guiding my thoughts, and to get me to this point of contentment!

Cheers,
Matt


(Abhinav Mathur) #519

Hi, any thoughts on following machine available at costco

It has 1070 8 GB graphics
32 GB RAM
and intel core i7

Will it be good to

  1. Participate with decent chance of getting ranking at kaggle for competitions like https://www.kaggle.com/c/cdiscount-image-classification-challenge
  2. Future proof for 2 years.

or should I go for https://www.costco.com/iBUYPOWER-C-i20-Gaming-Desktop---Intel-Core-i7-8700K---8GB-NVIDIA-GTX-1080-Graphics.product.100383125.html


(Shubham Singh Tomar) #520

Hi fellow,

I just wrote a blog post about how GPUs help with deep learning. The post also includes parts of Jeremy’s lesson 3 lecture. Could you guys please go through it and review?

Thanks :slight_smile:


(Aleksey Kazantsev) #521

Hi.
For me is best of devbox will be the server platform, they are much cheaper in our days (you can buy all stuff on ebay or https://natex.us/), for example:
Intel Xeon E5-2670 4 years old 2.6 8 cores 20M cach stepping 2 ~ 50$-150$ for each (ebay or natex)
DD3 2x 128GB 16x8GB ~ 500$ (ebay)
Matherboard? You have to find server matherboard for you process with PCIe x16, for example (natex) ~ 150-200$
2x geforce gtx 1080 ti or 2x geforce pascal titan 2
Server power supply, find youself ~ 50$ (natex)

Summary: 2 Xeon E5-2670 (16 core, 32 freads), 128GB from 800$ + Video ~ 600$

You can see all of this beauty here:
Dual Xeon E5-2670s Benchmarks - Build The BEST 16 Core USED Workstation
Dual Intel Xeon E5-2670 PC Server Build - 16 Cores, 32 hardware threads


(Xiaoru Li) #522

Hi, everyone. Apart from the hardware things, does anyone use his/her server remotely?

I’m considering to add a graphic card to my desktop in the dorm, and would love to access it remotely from the library or cafeteria. However my machine is behind a NAT (managed by the dorm/uni) + a WLAN router (mine), so there seems to be no simple way to ssh into it using my laptop from the library network (which I suppose is another unknown layer of routers/NATs). Does anyone have a similar network situation, and have come up with a solution? :grinning:

I’ve thought about joining the two machines together inside a VPN, and have looked into a virtual LAN gaming service Hamachi, but their linux client seems very buggy and useless. TeamViewer is also not a good option as what I really need is only Jupyter Notebook, not a remote desktop (laggy and different resolutions). :confused:

Xiaoru


(Aleksey Kazantsev) #523

Hello, what about router port forwarding? And all the problems go away ))


(Xiaoru Li) #524

I can configure my own router in my room, but that router itself is behind a NAT I don’t have access to. :confused: So I guess that won’t do.

(correct me if I’m wrong, I’m not very familiar with networks :grinning:)


(Aleksey Kazantsev) #525

Ask your provider (if this possible) to give you public IP adress.


(Cynosure) #526

It is also possible to buy static IP from many VPN providers. (google it).

I use a Fritz!Box (a DSL modem+router) which has its own VPN software (FRITZ!VPN) to connect to the router from anywhere in the world. I have not tested with desktop yet but with my mobile i can connect to my home network. It doesn’t has any recurring charges.


(Xiaoru Li) #527

Thanks for the tips guys! @KaaPex @cynosure

Unfortunately, I was just informed by my network admins that they couldn’t set up a port forwarding for me. Meanwhile I have done some googling on buying static IPs from VPN providers, but all of them seem to be much more expensive (all require subscriptions) than a low end VPS with static IPs.

Perhaps I’ll try setting up reverse SSH tunneling or a VPN server on a VPS after the holidays.

Thanks again and happy holidays!


(Cynosure) #528

What i meant is there are two options (which i know) to connect to your home network without getting a static IP from your ISP:

  1. get a static IP from a third party VPN Provider. e.g. check this. In this, you would get a static IP from this company assigned to your machine at home via VPN (no matter what IP your ISP will give you). I have not used or tested this but i suppose it can work. it has recurring charges but i think if it works it would still be cheaper than a static IP from the ISP. Details

  2. Get a Router called Fritz!Box. When you use it in your home network and setup its VPN what it basically does is it registers your home router (fritzbox) dynamic IP to FritzBox’s website/service with your user ID (and keeps updated if your ISP updates your dynamic IP) and when you run the Fritzbox’s VPN software on your remote PC from your office or a coffee shop its software goes to FritzBox’s website/service and gets (via your user ID) the dynamic IP of your home Fritzbox router from there (which was registered by your router) and creates a VPN Tunnel to your Fritzbox router at home, so you can use work anywhere like you’re at your home.
    Here are setup instructions. It doesn’t has any recurring charges except one time cost of FritzBox. Setup Instructions Video.
    Disclaimer, I have not tested it via Windows myself but I use its VPN App on my Android through which i can access my home network and landline phone records/mailbox. More info. For DL work, I will test and use its Windows VPN sometime in near future as well.

The router in second option is from a German company (i live in Germany) and is a popular router in Germany. I guess in the US or else where it is not so much popular but it should work anywhere IMO.


(Nikhil J) #529

After lurking on the thread for a lot of months and procrastinating on my build, I’ve finally decided to bite the bullet and finish the build.

Here’s my draft of a build:
https://pcpartpicker.com/user/seasponge/saved/h7B9WZ

I procured a AORUS 1080Ti on Massdrop a couple months ago, and it’s been sitting on my desk. I’ve just ordered a 8700k, just to future proof the build, as I would like to use the PC for gaming too occasionally.

I’m really vague about what to choose for a z370 MOBO, I feel like the MSI - Z370M GAMING PRO AC Micro ATX LGA1151 Motherboard ticks all the boxes for a deep learning build, but am not sure if I’m missing something.
My thought process on the pick was filtering for 64GB max ram, 2 x16 pcie slots, 4-5 stars on pcpartpicker and sorted by price ascending.

I’m currently looking at getting a 500 gb SSD, but am also glancing at NVMe drives. Might leave that upgrade for the future if the desire grows.

Please critique my build/offer suggestions, as to what I might be overlooking.


(Robert Salita) #531

A private server can be made public by using localhost.me (free) or ngrok (free and paid tiers). You can gain desktop access using VNC, Splashtop. You can file share using Dropbox, Google Cloud and others.


(Cynosure) #532

That sounds rather high end setup (for a DL beginner student) to me. If you are serious about DL and have the budget for it, go for it.


#533

I’m running a Windows 10 laptop (GeForce 1060, Core i7, 512GB SSD) and trying to set up a Windows 10 GPU server using this guide:

http://wiki.fast.ai/index.php/Local_install_(Windows_only:cpu)

I ran into a problem I can’t solve:
keras.json doesn’t seem to get created for some reason. I think I need to run import keras to get it to create the file.
However, running import keras gives me the following error:

“RuntimeError: To use MKL 2018 with Theano you MUST set “MKL_THREADING_LAYER=GNU” in your environement.” (sic)

I tried creating this system variable, but it made no difference.
Has anybody had any success getting a Windows 10 GPU server working? Or should I just give up and dual boot Ubuntu?


#534

Pekoto,

Two things to consider here,

First, it looks like you are trying to setup version1 of the course which uses Theano. Theano is dead. Version 2 of the course uses a different framework called Pytorch. Configuring Windows 10 for this is a bit easier and some steps can be found here. In short, you just download cuda 9.0.176? get cudnn v7, install latest anaconda and github desktop. I just set this up today on a laptop similar to yours and it appears to be working.

Second, you may find it easier in the end to go Ubuntu. I have a desktop that is dual booted with dual drives for Win10 and ubuntu. Once the OS is installed, the setup closely resembles the paperspace setup script.


#535

Got it, thanks for your tips. I’m going to try installing Ubuntu.


(Andrea de Luca) #536

I did. I got a working w10 box (i7, gtx1070) with keras/TF and fastai/pytorch in separate conda envs.

What did go awry with your attempts?

However: How to set up Windows 10 for fast.ai


(Omar Amin) #537

Thanks for the very informative thread, I’m building my own box now, can you please revise this part list and let me know if there’s anything wrong?

https://pcpartpicker.com/list/4qPV3b

I may add another GPU later, that’s why I’ve chosen the x99 motherboard that supports up to 2 gpus * 16x PCI lanes

I’m in short budget, I already have the titan gpu, I’m just creating the box around it, and I noticed the warning of the incompatibility for the cpu and the motherboard, but i think it’s not on this specific version of the CPU & motherboard


#538

i have the 6850 with an X99 msi motherboard, pcpartpicker gave me the same warning, but I did not have an issue.

I would recommend an NVME drive over the SSD you have.
I would also recommend a 1080ti over the Xp.

The HD recommendation is based on my setup. I think the nvme helps alot with file read/write operations. A 1080ti should be cheaper in theory than an Xp and may perform faster. Refer to this thread.