However, it will still charge even though you shutdown an instance
meaning you actually need to delete an instance if you are going somewhere else.
also it means that you need to re-instantiate (provision) the instance, install all the dependencies, download materials, download all the data that you need once you come back.
data is very important, and no one wants to waste of time on that after setting up once.
this is a terrible feature for DataCrunch I guess.
Sounds easy. Is GPU in use also or do we need to ”Play” with CUDA installation? It would be nice if datacrunch can make YouTube video from end to end : How to build one simple Notebook. Save It. And shutdown or delete so that patments stop. And last part how continue from this state.
We are currently developing ‘hibernation’ of instances; you will be able to save the state of your instance so you can deallocate your instance without losing your additional dependencies and data. It should be available in about 3 weeks.
Here you can find detailed instructions how to setup your instance with a Fastai image: https://course.fast.ai/start_datacrunch
When you order an instance with a Fastai image, it already has CUDA and other requirements installed to get started.
As mentioned above, we are working on storing the state of your instance so you don’t need to keep it on your account. Making a youtube video is a good tip, we will provide a video tutorial once it is possible to store your instance.
I didn’t find anything about the hibernation on your homepage.
Did the feature launch yet? If not, is there already an ETA?
I’m planning to start the course soon’ish and find your products extremely attractive. But having to delete and re-setup everything every weekend (when I have time to work on the course) would be a pretty big deterrent. A small storage fee for hibernating instances would be fine with me, though.
I’m asking here because I suspect other people may also be interested in this information.
We are not advertising it yet on the homepage but the feature is available.
Note that hibernation is currently in beta, we have not had problems since launch though.
@Wolf Thanks! Don’t hesitate to contact if you have any questions or recommendations.
Next up is releasing our API and expanding services towards easily deploying trained models.
@flowerkowski Thanks for the kind words! We try to optimize towards ease of use.
I cannot deploy a DataCrunch server because: “FastAi image is temporarily disabled”. When will I be able to? Is this maintenance-in-progress or an error?
I want to setup my local workstation with the same software configuration as I will have on datacrunch, so that I can rsync my data and changes to notebooks between them, with the goal of training models on datacrunch, with prepping datasets and low GPU stuff on my local machine.
The whole where do I start? thing for the fastai book/course needs curation.
To quote myself: “The reason you are confused is, because it is confusing.”
@DataCrunch just started using the server and the difference is night and day compared to collab. Thank for the discount.
Edit: Nevermind, found it inside the server overview.
Does it mean, now if I click shutdown everything is good and I won’t be charged money or do I have to do something special to activate this option. If there is a list of best users that have this access, I’d love to signup.
Everyone can use the feature, if you shut down the instance (using the shutdown button on your dashboard), you can click hibernate after it is shut down.
The hibernation function is not very practical if you have over 100GB of data at the moment, we are planning a complete overhaul in January.
The instance only switches to hibernation rate when the instance is actually hibernating, before that the hardware is still dedicated to the user. The time it takes depends on the amount of data that needs to be hibernated. For the average user/instance it is about $0.25.
Since it can take a long time currently for instances with a lot of data, we have disabled the function except for 1V100 instances.
We are transferring our servers to network storage instead of local storage, which will speed up these operations dramatically and won’t require the resources to stay dedicated to the user.