Kevin Burns, Virginia Polytechnic Institute, USA
Robert Lyerly, Virginia Polytechnic Institute, USA
Antonio Barbalace, Virginia Polytechnic Institute, USA
Binoy Ravindran, Virginia Polytechnic Institute, USA

Several decades ago (beginning in the early 1990s), there was a clear distinction between general purpose and real-time operating systems. This distinction has become increasingly blurred, in part due to the need to support time-sensitive multimedia services on a wide range of platforms (which requires an RTOS) and the need for system integration (i.e., non real-time services must be implemented on an RTOS, enabling its usage in a system/networked environment). In response to this trend, signif- icant attention has recently focused on optimizing Linux for embedded platforms, in particular, on improving its multimedia capabilities. This has also resulted in the development of the PREEMPT_RT patch (now integrated into the mainline) to reduce latencies and increase determinism in code execution. Commercial entities (e.g., LynxOS, Hawk) have introduced similar patches, and others (mostly universities) have attacked the same problem space with a dual kernel solution (e.g., RTLinux, RTAI, Xenomai). A further contribution from academia is the implementation of real-time scheduling policies in the Linux scheduler (e.g., LITMUS^RT , ChronOS, IRMOS).

In this paper, we analyze and compare PREEMPT_RT, SCHED_DEADLINE, IRMOS, LITMUS^RT , ChronOS, RTAI, and Xenomai to a vanilla Linux kernel, in order to obtain a better understanding of their usability and applicability in different contexts. Our metrics of interest include kernel overheads (e.g ., scheduling, task migration), schedulability (e.g., deadline satisfaction), and scalability to many cores. Our evaluation framework includes a port of a synthetic real-time application suite (that we developed) to the aforementioned real-time kernels. (A by-product of this endeavor was the extension of IBM’s librttest library to support the real-time API of a wider number of real-time kernels.) Our results show that the dual- kernel implementations are adept at handling low latency applications at low core counts, mainly because, they do not implement global multicore real-time scheduling policies. However, as the number of real-time tasks increase, these implementations begin to accumulate overhead due to expensive context switches (the comparisons run entirely in user-space).