RTAI: Real-Time Application Interface
From version 0.9, RTAI includes a proc interface which gives useful information on the current status of RTAI, including schedulers loaded; real-time tasks activity, priority and period; and FIFOs in use and their associated buffer sizes. The development of even more features is currently underway.
RTAI provides true support for symmetric multi-processing (SMP) architectures through its task and IRQ management.
By default, all tasks are assigned to run on any CPU (of a SMP platform). Each task, however, can be individually assigned to any CPU subset, or even to a single CPU. Additionally, it is possible to assign any real-time interrupt service to any specific CPU. Because the ability to force an interrupt to a specific CPU is not related to the SMP scheduler, RTAI retains the flexibility to perform these two operations independently.
These capabilities provide a method of statically optimizing the real-time application, if manual task distribution handles the task more efficiently than the automatic SMP load-distribution services of Linux.
Since real-time Linux tasks are implemented as loadable modules, they are, for all practical purposes, an integral part of the kernel. As such, these tasks are not bounded by the memory protection services of Linux, and they have the ability to overwrite system-critical areas of memory, bringing the system to an untimely halt. This limitation has been a large frustration to those of us who have erred during real-time task development.
RTAI's LXRT solves this problem by allowing the development of real-time tasks, using all of RTAI's hard real-time system calls from within the memory-protected space of Linux and under a “firm” real-time service. When the developer is satisfied with a task's performance within LXRT, the task is simply recompiled as a module and inserted into the kernel (along with the associated modules which provide RTAI's real-time features) to transition from firm to hard real-time.
LXRT's firm real-time service, similar to that offered by the Kansas University Real-Time (KURT) patch, provides soft real-time combined with fine-grained task scheduling. Performance under LXRT is quite good, yielding latencies not much greater than for a standard Linux system call leading to a task switch. Although this is very valuable as a development tool, we should not lose sight of the fact that RTAI's firm real-time implementation can prove especially useful for those tasks which don't require hard real-time, but yet are not quite satisfied with the scheduling performance of standard Linux.
RTAI implements a compliant subset of POSIX 1003.1.c through the use of a single loadable module. These calls support creation, deletion, attribute control and environment control for threads, mutexes and condition variables. The resultant POSIX support is similar to standard Linux threads, except that parent-child functions (which are not appropriate for real-time tasks, since all threads are considered to be part of a single process) and signal handling (which is currently in development) are not supported.
RTAI is now competitive from both a cost and performance perspective with the commercial RTOS currently available.
Since the performance of any RTOS system is determined by the performance of the RTOS itself, the performance of the hardware on which it is running, and the test procedure used to acquire the data, absolute performance figures are very difficult to quantify, often making comparisons difficult between fundamentally similar RTOSes. However, the data below was measured on, and is representative of, typical Pentium II 233MHz and 486 platforms.
For these performance characterizations, an early version of the RTHAL module was demonstrated running a timer at 125KHz (Pentium II, 233MHz) while simultaneously servicing Linux, which was working under a heavy load. During this demonstration, the average and maximum jitters about the periodic timer were 0<\#181>s and 13<\#181>s, respectively. This performance, combined with additional tests, can be summarized in this way:
Maximum periodic task iteration rate: 125KHz
Typical sampling task rate: 10KHz (Pentium 100)
Jitter at maximum task iteration rate: 0-13<\#181>s UP, 0-30<\#181>s SMP
One-shot interrupt integration rate: 30KHz (Pentium-class CPU), 10KHz (486-class CPU)
Context switching time: approximately 4<\#181>s