Right now, the bucket size is 32 and anything greater than it will be fell into the last bucket. It has been used widely in the Lustre code base, in particular the number of pending IO to disk is using this. This might be good enough for the HDD era but nowdays NVME SSD can support many more pending IOs. For example, it's not uncommon to see a NVME SSD device to support 10+ queue pairs with 1K queue size for each queue. 

This has caused issue to a recent performance investigation. I would propose to have a log2 based histogram.