You’re effectively using https://en.m.wikipedia.org/wiki/Little%27s_law to deduce component performance as service time. I’ve found the peer filtering mechanism to work as well. It can also work if the entire node is slow, depending upon characteristics of the workload.