I have been trying to run checkpointing (github cuda-checkpoint & CUDA Driver API) on Confidential Computing but I have been getting the error message “operation not supported”. Does Confidential Computing by design not let checkpointing or is it still in development?
It looks like it. We will take this question as a feature request, as it seems like at least one person is interested in it.
Given that checkpoint/restore ~= live migration, I guess NVIDIA Confidential Computing also does not support live migration by default?
This is correct, it is currently not supported. Live migration in confidential computing (with or without GPU) is a delicate task, as a malicious infrastructure owner could live migrate your data to a machine THEY own. It’s certainly a feature that is being investigated, however.