In multitask, preemptive real-time systems, the use of cache memories makes estimating the response time of tasks difficult, due to the dynamic, adaptive and non-predictable behaviour of cache memories. This work presents a comprehensive method for attaining predictability on the use of caches in real-time systems through the use of locking caches, which ensure cache contents will remain unchanged during the execution of each task. Nowadays, locking caches are present in several commercial processors. In order to select the contents to be locked in cache, a genetic algorithm has been developed. Experimental results indicate that this scheme has a high level of predictability, and that the performance loss is negligible for around 70% of the tasks.