We present a brokering service for the adaptive management of composite services. The goal of this broker is to dynamically adapt at runtime the composite service configuration, to fulfill the Service Level Agreements (SLAs) negotiated with different classes of requestors, despite variations of the operating environment. Differently from most of the current approaches, where the performance guarantees are characterized only in terms of bounds on average QoS metrics, we consider SLAs that also specify upper bounds on the percentile of the service response time, which are expected to better capture user perceived QoS. The adaptive composite service management is based on a service selection scheme that minimizes the service broker cost while guaranteeing the negotiated QoS to the different service classes. The optimal service selection is determined by means of a linear program that can be efficiently solved. As a result, the proposed approach is scalable and lends itself to an effi...