Making your own server

(Can Yi) #61

@lin how could i found the server name of my Nvidia, I have just reached the last step and can not figure out, google seems not help on this problem! thanks!

(Can Yi) #62

I have resolve this problem… hostname do not work for me, I used the IP to replace the server name, Thanks!!!

(D) #63

Welp, I got it up and running. Just did my first model, had to lower the batch size significantly.

My hardware is:
Ubuntu 16 (dual booting with windows)
i7-4770K (3.50 x 8)
GTX 760 (2 gigs)

I’m running faster than I did on the P2, so I’m happier.

I’m also thinking about running out and grabbing a new cards lol.

Last thing I havn’t gotten working is the SSH. Even after I give my server the pub key, I still cant connect.

(Rothrock) #64

Don’t know if anybody mention this site I’ve been using it to try and put together a machine for deep learning. Quite helpful.

(Christopher) #65

Very handy site, one nice thing is it will help validate compatibility for parts you pick, things like GPU being too big for the case you chose or having too small of a PSU for your equipment. I have been building PC’s since I was 13 years old, and I still will run my parts through it just to make sure there are no issues. Also very handy for getting best prices, but I generally like to shop from one vendor.

(Brendan Fortuner) #66

I assume you set up port forwarding on your home router? Have you tried checking if port 22 is open here: If it’s not, is the ssh server running: sudo service ssh status. You can also try password-based authentication to get started and move to pub key later.

(Jeremy Howard) #67

@brendan this is such a great writeup - I’ll be recommending all students in part 2 get a dedicated box, so I’ll send this link to everyone because it’s so helpful! :slight_smile:

(Anurag Goel) #68

Another vote for Here’s a link to my DL build:

PCPartPicker part list / Price breakdown by merchant

Type Item Price
CPU Intel Core i7-6850K 3.6GHz 6-Core Processor Purchased For $496.00
CPU Cooler Scythe Kotetsu 79.0 CFM CPU Cooler Purchased For $22.62
Thermal Compound Arctic Silver 5 High-Density Polysynthetic Silver 3.5g Thermal Paste Purchased For $7.88
Motherboard ASRock Fatal1ty X99 Professional Gaming i7 ATX LGA2011-3 Motherboard Purchased For $178.00
Memory G.Skill Ripjaws V Series 32GB (2 x 16GB) DDR4-2400 Memory Purchased For $154.00
Storage Crucial MX300 750GB 2.5" Solid State Drive Purchased For $87.49
Video Card GALAX GeForce GTX 1080 8GB EX OC Video Card (2-Way SLI) Purchased For $477.00
Video Card GALAX GeForce GTX 1080 8GB EX OC Video Card (2-Way SLI) Purchased For $477.00
Case Fractal Design Define R5 (White) ATX Mid Tower Case Purchased For $109.15
Power Supply EVGA SuperNOVA G2 1300W 80+ Gold Certified Fully-Modular ATX Power Supply Purchased For $143.81
Prices include shipping, taxes, rebates, and discounts
Total $2152.95
Generated by PCPartPicker 2017-02-19 15:55 EST-0500

(Anurag Goel) #69

Another forum I found useful to keep costs down as I was building my server:

(Christopher) #70

Really good community, very helpful provided you try to help yourself. They won’t spoon feed you but will help if you put the effort into researching.

(Rothrock) #71

I notice that it gives a compatibility warning that the X99 motherboard may need a BIOS update before you can use the Broadwell-E architecture. I’ve been getting those in my builds too. Any idea how to check if the mobo manufacturer has provided an update?

Also why did you choose Core i7 over Xeon E5? Just wondering. I’ve been going back and forth on which to choose.

(Anurag Goel) #72

I found that at this point pretty much all X99 mobos are shipping with firmware support for Broadwell-E. That warning was probably more useful in late 2015/early 2016.

Core i7 over Xeon was a minor cost optimization. The performance is similar and you don’t need to buy (typically) more expensive ECC RAM.

(Christopher) #73

Xeon is good when you want to go multiple processors, VFIO or a true server.

(D) #74

@brendan Yep, ports are working. Canyouseeme says its good to go, and I can ssh in and get the “password” section, but it seems like the folder the authorized_keys file is in needs different permissions or something.

But I can still nomachine in and since i’ve got the notebooks working (thanks for the password help) I’m up and running.

I think the next step is to upgrade to a 1060 or 1080. : ) Fun stuff.

(Sam Miller) #75

I built a new server this week for relatively cheap, $870 after rebates. The only thing really important is a good graphics card and like others have said the GTX 1070 seems to be the best performance for the dollar. I went with a MSI after reading about EVGA 1070’s & 1080’s exploding (catching fire). Apparently, they changed the reference design from NVIDIA and had some heat issues. The motherboard was Asrock z270 killer sli which has a lot of nice features like built in wifi and m2 slots, and of course gen3 pcie lanes 1@16x or 2@8x. The CPU, Intel G4560, is cheap but good enough that it won’t bottleneck anything and that’s all that matters. It has 16 pci lanes, only a dual core but intel put hyperthreading on kaby lake pentiums so 2 more virtual cores. The ram and case were whatever was cheapest at the time. Same with the power supply but if you change anything you need to calculate how much wattage you need. My build: CPU 54w + GPU 150w + MOBO & periph’s 85w = Total 289w is easily satisfied by a 430w PSU.

Here’s the complete parts list.

PCPartPicker part list

CPU: Intel Pentium G4560 3.5GHz Dual-Core Processor ($61.99 @ Jet)
Motherboard: ASRock Z270 Killer SLI/ac ATX LGA1151 Motherboard ($122.98 @ Newegg)
Memory: G.Skill TridentZ Series 16GB (2 x 8GB) DDR4-3200 Memory ($122.99 @ Newegg)
Storage: Intel 600p Series 256GB M.2-2280 Solid State Drive ($108.89 @ OutletPC)
Video Card: MSI GeForce GTX 1070 8GB Video Card ($394.99 @ SuperBiiz)
Case: Rosewill Stryker M ATX Mid Tower Case ($59.99 @ Newegg)
Power Supply: Corsair Builder 430W 80+ Bronze Certified ATX Power Supply ($52.77 @ B&H)
Total: $924.60
Prices include shipping, taxes, and discounts when available

(Sam Miller) #76

I get 234s on the lesson 1 benchmark that everyone is using. @dradientgescent is still the champ w/229s but I may try and optimize further.

(Christopher) #77

Are you running Windows?
If so, run Passmark Performance Test for 3D GPU Tests.

I got a score of 12981 on my 1070 not overclocked.

If you are on Linux, you can use GPUTest which is a different test but it is cross platform.

You should be able to get better, but different cards of the same model have different clock speeds and can perform differently. I am using the MSI 1070 Gaming X card which is one of the best 1070 cards.

(Sam Miller) #78

I’m running Ubuntu 16.04. It consumes less resources than Windows, disk and memory, so I can set [lib] cnmem = 0.95. I haven’t run any GPU tests or delved into nvidia-smi persistence and clock setting yet. I don’t know if I will or not but I was just going to chalk the 5s difference up to variance. It could be hardware variance, since my card is an MSI 1070 Armor, could also be some software variance.

(Christopher) #79

I thought you were the one with the 1080. Yeah I just likely have higher stock clocks. Each manufacturer overclocks the cards differently, then you can further overclock it on your own.

Linux will allow you to use much more ram, I am stuck to 80% (haven’t tested 85, but 90% does fail) with Windows. Performance-wise, it does well.

(Yashar) #80

I have a server with two GPUs. I installed all the required libraries and drivers. If I login to the server and run ipython and then run ‘import bcvolz’ then it works fine, but if I do it via jupyter notebook ( via port forwarding) I get the error
ImportError Traceback (most recent call last)
in ()
----> 1 import utils;
2 reload(utils)
3 from utils import plots

/raid1/home/yashar/courses/deeplearning1/nbs/ in ()
20 from scipy.ndimage import imread
21 from sklearn.metrics import confusion_matrix
—> 22 import bcolz
23 from sklearn.preprocessing import OneHotEncoder
24 from sklearn.manifold import TSNE

ImportError: No module named bcolz

If I do conda list I see bcolz as an installed package.
any suggestion?