Closed willprice closed 3 years ago
We've built a CUDA image using this configuration, but finalising the AMI takes a long time and so does booting up the instance. I'm unsure whether this serialises the whole 40GB or whether it snapshots at the file level. If it does snapshot the whole 40GB then we should decrease the size.
I think it snapshots the whole 40 GB. Perhaps 20 is enough?
I've just bumped it down to 15GB and that was enough for CUDA and the nvidia drivers, yet it still seems it takes a very long time to finalise the AMI. I haven't tried launching a node yet with the new AMI, but once it completes I will. I suspect they compress the image so that unused space doesn't actually equate to additional bandwidth needed to transfer the image.
Still, I think 20GB is probably a more reasonable size. Are you OK to make the change @milliams or do you want me to push an update?
I'll make the change, no problem.
By default it seems the AWS builder has a cap at 8GB of file storage. Quickly this becomes an issue when installing large toolchains like CUDA. This change gives the builder a 40GB block store which is snapshotted at the end of the build process to form the AMI.