why too mach requests will increase IO latency
 

I am learning block layer. I found a patch in this link.https://patchwork.kernel.org/project...ei@redhat.com/

He mentioned in it that "the default nr_requests for scheduler is too big for small queuedepth devices, and latency is increased much."

I don't understand why a too larger nr_requests can let to bad effect.
If there are few requests, wouldn't it quickly exhaust the tags and cause IO blocking?

The bad cases for I/O frequently have one greedy process launching a lot of I/O and other victim processes launching few. What happens to latency for the victims if you let the greedy process fill the queue instead of blocking?