Situation: Several millions of directories (representing class labels) with small images inside
On AWS:
Because of the IO ops count limit per second, iterating over the dataset takes forever and IO performance is the bottleneck. I found organizing data using pytables really helped and experimented with chunk sizes, etc.
Question:
How does performance look like on a regular box, with local SSD / HDD? Meaning, how big is the penalty hit if reading small files from a lot of directories vs reading small files from a single directory? Can I expect a significant improvement if I read in data from HDF5 files?
Probably the answer here is go and see if performance is adequate starting with the simplest solution, but was just curious if maybe anyone has any experience on this / any theoretical knowledge that they would be willing to share 
In general - does amount of files / reading from various directories visibly impact reading in speed on an ext4 filesystem?