Measuring Performance of Soft Real-Time Tasks on Multi-core Systems

Abstract: Multi-core platforms are well established, and they are slowly moving into the area of embedded and real-time systems. Nowadays to take advantage of multi-core systems in terms of throughput, soft real-time applications are run together with general purpose applications under an operating system such as Linux. But due to shared hardware resources in multi-core architectures, it is likely that these applications will interfere and compete with each other. This can cause slower response times for soft real-time tasks. In order to investigate this problem, a number of memory intensive and computation intensive soft real-time tasks were co-scheduled with Linux SMP running a general purpose task on it. For performance measurement a test environment is created that uses hardware registers to count core, level-2 cache and memory bus events at run-time instead of having a simulator tool. In particular events related to L1 and L2 instruction and data cache, memory bus utilization and difference in response times of soft real-time tasks are measured. After completing this research, we can say that it is only possible for soft real-time applications to co-exist with Linux SMP cores running general purpose applications without major performance degradation, if the task on Linux is not much memory intensive. If it is memory intensive then there is a trade-off between number of cores running general purpose applications and the amount of tolerance an embedded system can have in response times.