Draining rate drops and becomes choppy when more than 5% cache miss happens

3 node cluster, 4 cores, 8G RAM for each node. The disk is 30G HHD.
Create a default bucket with 3.5G RAM.
Initial Load 35M items to make the cluster into 75% resident ratio and when access the data the workload definition is:
create:10, update:30, get:50, delete:10,ops per sec:30000

When many background fetches happen, cache miss around 5%~10%, contention to the disk write queue between threads make the draining rate drop significantly and choppy.