Slowest Server Emulation Syndrome: A Hidden Cost of Load Imbalance

We model a certain class of distributed "shared nothing" service in the presence of heavily loaded clients, and show, through mathematical models and simulation, that the system throughput can drop to the point that all servers process tasks at the rate of the slowest server. We also run experiments on an instrumented deployment of a commercially available database and find this same result. We present formulas for equilibrium values and for the time constants that characterize the approach to equilibrium.