TITLE: Estimating the confidence in the QoS guarantees of internet services AUTHOR: Tudor Dumitras ABSTRACT It is essential for internet service providers to articulate a set of measurable QoS guarantees that are highly relevant for the client and that can be satisfactorily upheld by the provider with some level of confidence. However, using service-level agreements based on hard guarantees on client-required QoS metrics (e.g., end-to-end latency, throughput, availability) might be risky for service providers because this behavior depends on unpredictable conditions in the environment (e.g., network latencies, resource unavailability), as well as load and input data. This might be further complicated by the fact that QoS metrics themselves tend to exhibit non-linear, potentially conflicting relationships with each other. This paper attempts to evaluate experimentally the level of confidence in the QoS guarantees of a Fault-Tolerant CORBA system called MEAD. The data suggests that by guaranteeing statistical confidence intervals rather than hard limits for the QoS metrics (e.g., "The response time of the service shall not exceed X seconds for more than 1% of the requests"), the service providers can implement and enforce the guarantees more easily because the erratic behavior is not accounted for. Such statistical guarantees are also relevant for the clients because they accurately capture the average behavior of the service.