Methods, systems, and computer-readable media for determining the maximum throughput of a service are disclosed. A first sequence of load tests is initiated for a service host. Individual ones of the load tests comprise determining a respective throughput at the service host for a respective number of concurrent connections to the service host. The number of concurrent connections increases nonlinearly in at least a portion of the first sequence of load tests. The first sequence of load tests is discontinued when the throughput is not increased by a threshold from one of the load tests to the next. An estimated maximum throughput for the service host is determined based at least in part on the first sequence of load tests. The estimated maximum throughput corresponds to a particular number of concurrent connections to the service host.