New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Computing Server for 1 week
I need a very powerful dedicated server to use in the next week I have a budget of £20 for 7 days usage I need as much computing power for a machine learning project. It needs at least 10GB SSD storage 100mbit/s or less bandwidth and be as powerful as possible. Please post the results of running: python -m timeit '"-".join(str(n) for n in range(100))'
on the command line.
Also please do not post if it is any greater than 15 seconds as I already have a server that can do that.
Comments
Wholesaleinternet with hourly servers and OVH comes to mind.
you mean if its >15 seconds its slow?
Thanks but I posted on here to get the results before purchasing. I also need it for a week only.
yes
?
OVH Weekly servers
I know its not SSD https://www.ovh.co.uk/dedicated_servers/one-week-rental.xml
I was hoping I could get cheaper with a hell of a lot less storage, more cpu and less bandwidth!
Vultr dedicated hourly price is right up your alley. Can't test for you right now but figured I'd bring it up.
also scaleway.com maybe try hourly options first.
https://my.wholesaleinternet.net/?clientapp=launch_hourly_servers
I7-3700 120gb SSD + 2TB Preconfigured $0.07/hour
You have to log in
FYI, you can create an account without buying an actual server.
http://i.imgur.com/QhmVGc0.png
Would you be able to get results of the python command?
Ummm I'll spin up the i7 if you want
Edit: WSI rejected card
It costs 0.07, don't tell me thats too expensive....
The i7 did
Want me to test any others?
On my Hetzner i7-3770, that benchmark took about 5.7 seconds. Obviously it's a single core test on a 4 core machine. If it helps, there are hourly GPU instances available at paperspace.com. I don't know if that would be more cost effective for your project. They are much more expensive than normal dedis.
Wowzers how much is that to run every hour? I also am not sure how to run sklearn commands in multiple threads!
Is that the 0.07 per hour? Well my budget is £20 for about 4 days so that is anything under $0.27 per hour. Would you be able to find me the best for that price?
It's only available by the month, but machines with that cpu start around 30 euro/month inc vat:
https://robot.your-server.de/order/market/country/GB?hdsize=3000
That one has 16gb ram, 2x 3TB hard drives, but no SSD.
It's the same cpu as the WSI one for 7 cents/hour = around 50 usd/month. Disk setup is a little different.
If you do this type of thing much I think it's worth having a monthly one.
You may want to profile your sklearn code on your own multiprocessor machine. Most of the underlying linear algebra ops are already parallelized by the version of BLAS linked. Many higher-level machine learning functions are parallelized, too. And of course a lot of tasks in machine learning are embarrassingly parallel, e.g., cross-validation, random forests and other ensemble methods, etc.
If you are in college, subscribe to Continuum to get Python linked with Intel's Math Kernel Libraries (MKL). MKL will use multiple cores automatically where possible.
Alternately, you could compile Numpy/Scipy with Intel MKL or OpenBLAS. Numpy is the building block for most ML packages in Python, including SKLearn.
GPU is the real game-changer though. I don't know if SKLearn supports GPU directly, but you could link in NVidia's CUDA BLAS, for instance, to re-route BLAS requests to the GPU.
Tensorflow uses gpus directly. No idea if it does the stuff you need.
weird, it never gets under 15.6 here, yet I have to admit that I'm running virtualbox instance.
This command
python -m timeit '"-".join(str(n) for n in range(100))'
is single threaded, while most modern processor are multiple core. Hopefully your workload is parallelized and can use multiple threads. A better test might be to use sysbench (sysbench --test=cpu --cpu-max-prime=20000 run
) for a single threaded run, and sysbench (--test=cpu --cpu-max-prime=100000 --num-threads=<cpu cores>
) for multiple cpu machines.I can let you use 2 x E5-2670's for 10c / hour / each. (E5-2670, 128GB ram).
Woah, you mean a dual E5-2670 box for 0.10/hour? I'd be interested in that: do you have them generally available?
Thanks!
@willie Only as batch compute nodes (think scientific data, or computationally expensive tasks). Nothing interactive (ie, web hosting, VPS, spam, scraping, etc).
I'll offer $7/mo if I can use them for idling
What am I missing ? Vultr/DO/Linode ?