I've been using AI Training for several big trainings and every time the training time is abnormaly long (And billed accordingly). It takes 30 days for a training that should take not longer than 5 days.
The Job Monitoring interface shows that both CPU and GPU are idle most of the time.
I use a custom image based on python:3.9 (which is based on debian buster).
Data are located in a mounted object storage RW:cache. (Note that the problem was the same without the cache). Outputs are stored in the same storage.
I suspect an IO or journalisation problem, or a problem related to the use/sync of object storage but i cannot inquire it as IO monitoring is not available.
Am i doing something wrong ?