Details
Description
During the perf testing, we found that when we scale up to 20k nodes, the backlog size is often fullĀ and cause allocation threads to wait. Hence, it's better to make this configurable so user can tune the perf with respect to different size of nodes.