Service providers face the challenge of meeting service-level agreements (SLAs) under uncertainty on the application actual performance. The performance heavily depends on the characteristics of the hardware on which the application is deployed, on the application architecture, as well as on the user workload. Although many models have been proposed for the performance prediction of software applications, most of them focus on average measures, e.g., mean response times. However, SLAs are often set in terms of percentiles, such that a given portion of requests receive a predefined service level, e.g., 95% of the requests should face a response time of at most 10 ms. To enable the effective prediction of this type of measures, in this paper we use fluid models for the computation of the probability distribution of performance measures relevant for SLAs. Our models are automatically built from a Palladio Component Model (PCM) instance, thus allowing the SLA assessment directly from the PCM specification. This provides an scalable alternative for SLA assessment within the PCM framework, as currently this is supported by means of simulation only.