SMP and Embedded Real Time
There is hard real time, which offers unconditional guarantees, and there is soft real time, which does not. What else do you need to know?
As it turns out, quite a bit. There are at least four different definitions of hard real time. Needless to say, it is important to understand which one your users have in mind.
In one definition of hard real time, the system always must meet its deadlines. However, if you show me a hard real-time system, I will show you the hammer that will cause it to miss its deadlines, as shown in Figure 2.
Of course, this is unfair. After all, we cannot blame software for hardware failures that it did not cause. Therefore, in another definition of hard real time, the system always must meet its deadlines, but only in absence of hardware failure. This divide-and-conquer approach can simplify things, but, as shown in Figure 3, it is not sufficient at the system level. Nonetheless, this definition can be useful given restrictions on the environment, including:
Memory-system overhead due to DMA.
Memory-error rate in ECC-protected systems.
Packet-loss rate in systems requiring networking.
If these restrictions are violated, the system is permitted to miss its deadlines. For example, if a hyperactive interrupt system delivered an interrupt after each instruction, the appropriate action might be to replace the broken hardware rather than code around it. After all, if this degenerate situation must be accounted for, the latencies will likely become uselessly long. Alternatively, “diamond hard” real-time operating systems and applications might run with interrupts disabled, giving up compatibility with off-the-shelf software in order to gain additional robustness in face of hardware failure.
In yet another definition of hard real time, the system is allowed to miss its deadline, but only if it announces its failure within the deadline specified. This sort of definition can be useful in data-fusion applications. For example, a system might have a high-precision location sensor with unpredictable processing overhead as well as a rough-and-ready location sensor with deterministic processing overhead. A reasonable hard real-time strategy would be to give the high-precision sensor a fixed amount of time to get its job done, and if it fails to do so, abort its calculation, relying instead on the rough-and-ready sensor. However, one (useless) way to meet the letter of this definition would be to announce failure unconditionally, as illustrated by Figure 4. Clearly, a useful system almost always would complete its work in time (and this observation applies to soft real-time systems as well).
Finally, some define hard real time with a test suite: a system passing the test is labeled hard real time. Purists might object, demanding instead a mathematical proof. However, given that proofs can be subject to error, especially for today's complex systems, a test suite can be an excellent additional proof point. I certainly do not wish to put my life at the mercy of untested software!
This is not to say that hard real time is undefined or useless. Instead, “hard real time” is the start of a conversation rather than a complete requirement. You should ask the following questions:
Which operations must provide hard real-time response? (For example, I have yet to run across a requirement for real-time filesystem unmounting.)
What is the deadline? A ten-millisecond deadline is one thing; a one-microsecond deadline is quite another.
What is to happen in case of hardware failure?
What is the required probability of meeting that deadline? (For hard real time, this will be 100%.)
What degradation of non-real-time performance, throughput and scalability can be tolerated?
One piece of good news is that real-time deadlines that once required extreme measures are now easily met with off-the-shelf hardware and open-source software, courtesy of Moore's Law.
But, what if your real-time application is to run on an embedded multicore/multithreaded system? How can you deal with both real-time deadlines and parallel programming?