Systems and methods are described for allocating requests to implement new workloads within a heterogenous fleet. The fleet can include various sub-fleets, each corresponding to a set of computing devices having a given configuration of computing resources. A routing device can calculate n-dimensional decision surfaces that map expected resource usage associated with an incoming request to probabilities to route the request to each sub-fleet. The decision surfaces can be calculated to maximize cost-weighted headroom across the sub-fleets, with headroom on each sub-fleet reflecting a geometrical dissimilarity in a shape of load on the sub-fleet and a shape of resources available on the sub-fleet. By comparing the expected resource usage associated with the incoming request to the decision surfaces, the device can determine a sub-fleet to which to route the requests.