How PFC, ECN and DCQCN actually behave once you build the lab and measure it yourself. Every number here came off the local simulator, and every equation is checked against that measurement rather than quoted from a datasheet.
The fabric
The first thing to get straight is where the congestion actually is, because it is not where most people look. The trunk between the two switches is fine. The problem is the last hop into the receiver, where two 40 G senders are pointed at a single 40 G port.
Three flows run for 100 ms. h1 and h2 both send to h3, which is the 2:1 incast. Separately, h0 sends to h4 along a path that shares only the trunk with the other two. h0 is the victim flow. It takes no part in the congestion but, as Task 1 shows, it pays for it anyway.