Performance testing of multi-user software systems is typically performed by emulating activities of multiple users working with the system simultaneously. These virtual users collect performance measures like response time and correctness while loading the system during the test. During a typical load scenario number of simultaneous virtual users gradually increases from one to some intended maximum, allowing to find out statistical relation between number of simultaneous users and performance measures. However having fixed load scenario has several drawbacks. Usually the test has to be repeated iteratively with adjusted load scenarios, taking into account results acquired from previous test iterations. As load scenarios have to be adjusted manually and repetitions of the test may take considerable time, this approach seems to be ineffective. This paper presents an alternative approach, where load scenario is adjusted automatically during the test, taking into account performance measures already gathered in the same test. Drawbacks of fixed load scenarios are analysed and methods of eliminating them by automatic adjustment are described. A case study showing implementation of the method and results obtained by applying it to performance testing of a real system is provided.