Efficient verification of performability guarantees

The performability of a system is usually evaluated as the mean of a performance measure, aggregated over all failure scenarios (both tolerated and not) through a Markov reward process. This metric can gloss over performance deficiencies in important portions of the failure scenario space, and is also expensive to compute because the number of possible failure scenarios can be very large. We argue that a multi-part performability specification better represents the requirements of system designers and present a fast algorithm for determining whether a given system meets such specifications.