We consider the efficiency of dynamic resource pooling and allocation in a cloud computing system offering infrastructure-as-a-service (IaaS). We assume that the demand for service computing by virtual machines (VMs) follows a Poisson load pattern and that the response times of the provided computing services can be classified into several service categories that are governed by exponential service time patterns. A hierarchical, dynamic, class-dependent balancing policy based on a least-loading scheme is applied to provide a uniform utilization among the servers. It is derived from cascaded mutual overflow routing using information on the utilization of VM clusters of similar type on adjacent servers within this resource pool. Regarding the allocation of virtual machines of these different service types to the user demand by a pool of physical servers, we derive a Markovian loss model with adaptive routing induced by cascaded mutual overflow as effective, state-dependent load balancing policy. We determine its basic performance characteristics applying a Markovian fixed-point model. Based on the latter we gain insight on the power of the proposed dynamic load balancing policy among service classes.