While real-time applications are becoming more and more concurrent and complex, the drive toward multicore systems raises new challenges related to the parallelization of such performance-critical applications. Transactional memory is an attractive concept for expressing parallelism for programming multicore systems as it avoids the problems of lock-based methods and eases programming. However, it has not yet been exploited for real-time systems. In this paper, we propose the first real-time directed case study of software transactional memory. In particular, our goal is to identify the origin of the variation of the worst-case execution times (WCET) of transactions in memory. Based on a real implementation, we show through various experiments that for soft real-time, transactions rollback times are not the main cause of execution times variation. A good memory allocator must also be provided in order to suitably bound the WCETs of transactions into software transactional memory.