On the Influence of Shared Memory Contention in Real-time Multicore Applications

The continuous evolution of processor technology has allowed the utilization of multicore architectures in the embedded system domain. A major part of embedded systems, however, are inherently real-time (soft and hard) and the use of multicores in this domain is not straightforward due to their unpredictability in bounding worst-case execution scenarios. One of the main factors for unpredictability is the coherence through memory hierarchy. This paper characterizes the influence of contention for shared data memory in the context of embedded real-time applications. By using a benchmark, we have measured the impact of excessive shared memory invalidations on five processors with three different cache-coherence protocols (MESI, MOESI, and MESIF) and two memory organizations (UMA and ccNUMA). Results have shown that the execution time of an application is affected by the contention for shared memory (up to 3.8 times slower). We also provide an analysis on Hardware Performance Counters (HPCs) and propose to use them in order to monitor and detect excessive memory invalidations at run-time.