On the LambdaLabs cluster they don't have persistent storage. It also takes a very long time to transfer pre-recorded data.
The solution: Speed up the recording process. Currently only one environment is recorded at a time, which is very slow, taking many hours to record a dataset of a sufficient size (at least 30k episodes). It should be easy to parallelise this and speed it up several fold, which means we can just re-record the data when we spin up an instance and it won't take too long.
If this enhancement saves us half a day of computation on an instance, then that saves us >$15. I expect it'll save us several times more than that because we're probably going to be loading up several instances.
On the LambdaLabs cluster they don't have persistent storage. It also takes a very long time to transfer pre-recorded data.
The solution: Speed up the recording process. Currently only one environment is recorded at a time, which is very slow, taking many hours to record a dataset of a sufficient size (at least 30k episodes). It should be easy to parallelise this and speed it up several fold, which means we can just re-record the data when we spin up an instance and it won't take too long.
If this enhancement saves us half a day of computation on an instance, then that saves us >$15. I expect it'll save us several times more than that because we're probably going to be loading up several instances.