Real-Time and Performance Improvements in the 2.6 Linux Kernel
Three projects for real-time support under Linux currently are active: the dual-kernel approach used by the RTAI Project and by products from embedded Linux vendors such as FSMLabs; a real-time Linux project hosted by MontaVista, an embedded Linux vendor; and freely available preemptibility and real-time work being done by Ingo Molnar and others, which is discussed openly on the Linux Kernel mailing list and which the MontaVista Project depends upon. In addition to these core kernel projects, other supporting projects, such as robust mutexes and high-resolution timers, add specific enhancements that contribute to a complete solution for real-time applications under Linux.
The dual-kernel approach to real time is an interesting approach to real-time applications under Linux. In this approach, the system actually runs a small real-time kernel that is not Linux, but which runs Linux as its lowest-priority process. Real-time applications specifically written for the non-Linux kernel using an associated real-time application interface execute within that kernel at a higher priority than Linux or any Linux application, but they can exchange data with Linux applications. Although this is a technically interesting approach to running real-time applications while using a Linux system, it avoids the question of general Linux kernel preemption and performance improvements. Therefore, it is not all that interesting from a core Linux development perspective.
MontaVista's Project to further real-time Linux leverages much of the existing work being done by Ingo Molar and other Linux kernel contributors, but it includes some additional prototype patches available only on the MontaVista Web site. The current patches available there are for a release candidate for the 2.6.9 Linux kernel (rc4). Therefore, they did not apply cleanly against official drops of the Linux kernel, which is moving toward 2.6.11 at the time of this writing. As such, the results from this project could not be included in this article.
The real-time, scheduling and preemptibility work being done by Ingo Molnar, the author of the O(1) Linux scheduler, and others has a significant amount of momentum, enhances the core Linux kernel and provides up-to-date patches designed to improve system scheduling, minimize latency and further increase preemptibility.
These patches have an enthusiastic following in the Linux community and include contributions from developers at many different groups and organizations, including Raytheon; embedded Linux vendors such as TimeSys; and from the Linux audio community. These patches provide capabilities such as heightening system responsiveness and minimizing the impact of interrupts by dividing interrupt handling into two parts, an immediate hardware response and a schedulable interrupt processing component. As the name suggests, interrupts are requests that require immediate system attention. Schedulable interrupt handling, more commonly known as soft IRQs, minimizes the impact of interrupts on general system responsiveness and performance.
The illustrations in the next section focus on comparing benchmark results from various vanilla Linux kernels against those obtained by applying the real-time, scheduling and preemptibility patches done by Ingo Molnar and others. These patches are up to date and provide complete, core Linux kernel enhancements that can provide direct benefits to Linux users who want to incorporate them into their projects and products.
In 2002, the Linux Journal Web site published an article titled “Realfeel Test of the Preemptible Kernel Patch”, written by Andrew Webber. This article used an open benchmark called Realfeel, written by Mark Hahn, to compare preemption and responsiveness between the standard Linux 2.4 kernel and a kernel against which Robert Love's preemption patch had been applied. Realfeel issues periodic interrupts and compares the time needed for the computer to respond to these interrupts and the projected optimal response time of the system. The difference between the time when an interrupt request is issued and when it is handled is an example of latency, and the difference between predicted and actual latency is known as jitter. Jitter commonly is used as a way of measuring and comparing system responsiveness.
This article uses the same benchmark application as Webber's article but imposes substantially more load on the system when measuring results. This is a technique commonly applied when benchmarking real-time operating systems, because even non-real-time operating systems may exhibit low latencies in unloaded or lightly loaded situations. The graphics in the next sections also present the results differently to make it easier to visualize and compare the differences between latency on various Linux kernels.