Heechul Yun, University of Kansas

Using multicore processors for real-time systems is challenging because of high performance variations due to contention in the shared hardware resources – e.g., shared last level cache (LLC), hardware prefetcher, DRAM controller, and DRAM.

In our previous work, we presented a kernel level approach, called MemGuard, which minimizes such performance variations by providing per-core memory (DRAM) bandwidth guarantee. It was evaluated using the SPEC2006 benchmark suite on a restricted hardware configuration in which cores use neither shared caches nor hardware prefetchers. However, because it used the SPEC2006 suite (non-real-time applications) and the average instruction-per-cycle (IPC) metric, its impacts on real-time applications were not shown in the work. Also, its effectiveness in the presence of other shared hardware resources---a shared LLC and hardware prefetchers – was not shown, as it was mainly evaluated on a platform with per-core private cache and no hardware prefetchers (manually disabled).

In this work, we present a case study of a real-time system (which runs two applications: a periodic real-time application with a given deadline and a non-real-time application) and demonstrates how MemGuard can be configured to deliver required real-time performance without significantly sacrificing the CPU utilization in three different hardware configurations – including one that shares a LLC and uses hardware prefetchers.

In this study, contention in shared resources cause 31 ~ 108% WCET increases of the real-time application, depending on hardware configurations, resulting 28 ~ 62% deadline violations. Using MemGuard, we can configure the system to eliminate all deadline violations with only causing 14~29% CPU utilization reduction of the non-real-time application. The results suggest that MemGuard can provide controllable real-time performance at a reasonable cost. Moreover, it is shown to be effective even in the presence of the shared LLC and/or hardware prefetchers, albeit to a lesser extent.