The ATLAS experiment at the LHC can record about 1 kHz of physics collisions, out of the LHC design bunch crossing rate of 40 MHz. To achieve a high selection efficiency for rare physics events while reducing the significant background rate, a two-level trigger system is used. The event selection is based on physics signatures, such as the presence of energetic leptons, photons, jets or missing energy. In addition, the trigger system can exploit algorithms using topological information and multivariate methods to carry out the filtering for the many physics analyses pursued by the ATLAS collaboration. In Run 2, around 1,500 individual selection paths, the trigger chains, were used for data taking with specified rate and bandwidth assignments. We will give an overview of the Run 2 trigger menu and its performance, which supports a broad physics program. We present the tools that allow us to predict and optimise the trigger rates and CPU consumption for the anticipated LHC luminosity. They are essential components to react to the changing LHC conditions and data taking scenarios. As an outlook to the upcoming ATLAS data-taking period in Run 3 from 2022 onward, we present the design principles and ongoing implementation of the new trigger software in a multi-threaded framework AthenaMT together with some outlook to the expected performance improvements.