Memory among Gpus isnot distributed evenly and it is crazy slow

When I am using my model if I go 1 gpu and batch 1 it runs okay,
but when i use dataparallel and use 4 gpus and batch size 4, it becomes crazy slow and also i notirce that the memory between the 4 gpus is distributed unequally (the first gpu uses way more).
can anyone please help me how should i fix this issue or how should i look for the problem? :cry:


Does this have anything to do with the fact that my data is really big (7TB) and therefore dataloader become slower in when i use dataparallel??