ATLAS is one of the two general-purpose detectors at the Large Hadron Collider (LHC). Using fast reconstruction algorithms, its trigger system needs to efficiently reject a huge rate of background events and still select potentially interesting ones with good efficiency. After a first processing level using custom electronics, the trigger selection is made by software running on two processor farms, designed to have a total of around 2000 multicore machines in a system called High-Level Trigger (HLT). The 2008 LHC startup and short single-beam run provided a ??stress test?? of the trigger. Following this period, ATLAS continued to collect cosmic-ray events for detector alignment and calibration purposes. These running periods allowed testing the HLT in different running conditions. This paper focuses on the experience gained in running the trigger in the fast-changing environment of the detector commissioning.