The question of what defines a real-time operating system is constantly discussed in our group. The definition of real-time behavior (“holds a deadline”) is common and widely accepted in the scientific community. However, I made the mistake of applying this definition also the OS scope.

One could argue that NT is no a real-time system, since it cannot provide guaranteed timing behavior, especially with the DPC / APC and interrupt dispatch level concept. An interrupt service routine can schedule deferred procedure calls (DPC’s). NT has the concept of interrupt dispatch levels, where the OS scheduler runs on the lowest level, and ISR’s run on the higher levels. DPC’s are executed on the level above the scheduler level, so if the DPC is executed, nothing else runs in the system. Since NT executes all DPC’s until the queue is emtpy, and since there is only one queue, it can happen that a malicious driver implementation schedules DPC’s which take a non-deterministic amount of time.

(BTW: You might not now that when I talk about Windows NT, it means also Windows 2000 / XP / whatever comes in the future. The operating system kernel remained more or less the same in the last years. Read the book from Solomon and Russinovich if you need more details about changes between the versions.)

However, the DPC problem mainly relates to the design of drivers, applications and hardware behavior. In fact, you cannot the blame the operating system itself for being RT-capable or not. With well behaving hardware, good drivers (with deterministic DPC timing behavior) and an according application it would be possible to implement a RT-application with NT.

We see that with a timing-only definition, everything could be a RT operating system. For this reason, our RT expert (and Wikipedia) define a RT operating system as set of facilities which enables the development of RT applications. This includes fixed priority scheduling, according API’s, high resolution timers, memory pinning, and so on. This set of features is not centrally defined. The QNX authors define one neccessary feature set for RT operating systems, POSIX another one. Therefore, the operating system authors define if it provides facilities for the development of RT applications, and if it therefore can be named as RT operating system. The users decide if they agree that the feature set is complete enough for the development of their RT application.

Finally for me it comes out that Windows NT is not a real-time operating system, since it was not intended to be one. The good thing is that Microsoft somehow agrees to this kind of definition:

The funny question is if Linux could be named now as real-time operating system, since it supports the POSIX real-time API’s out of the box. The answer is simple: The operating system authors do not define a real-time system with the POSIX RT feature set only, and therefore it is not one.

I saw a live demonstration of this tool during a presentation from MSR Cambridge. This tool rocks:

http://research.microsoft.com/projects/groupshot/