Panel Session: Predictability in Active Real-Time/Temporal Databases

Abstract

Jack Stankovic - I thought I would try a little bit of an analogy talking about what is predictability. We look at, say, a real-time system and what has happened in the past; there are these approaches where the system had to be predictable, for safety critical systems. So people actually built their own special purpose real-time operating system to have complete control over it so they can make it predictable. Then there were lots of demands for using off-the-shelf things, so people tried to use commercial off-the-shelf operating systems like UNIX, and would it be predictable? No, it was terrible, it didn’t really meet any real-time constraint - it was hard to say what the system would do.

And then there’s this conversion that we take a UNIX and make it a real-time UNIX - try to undo the features that are bad for real-time performance and then do we get a degree of predictability? Well, it’s somewhere in between, in fact, you get a lot better than just using off-the-shelf technology, but it still leaves something to be desired, compared to using a special purpose operating system. It is essentially a trade-off that we are giving up some degree of predictability for the benefit of using a commercial kernel that people are used to.

When we look at the database, we might see almost a similar pattern, I think. In real-time databases, if we want predictability and the people have built their own small databases for embedded systems, telecommunications, or air-traffic control, a lot of time these databases are main-memory resident, their transactions are pre-compiled, so they are really trying to be very precise about everything, building it up from scratch.

These are transcripts of statements of the panel members, edited only for completeness, brevity, and clarity.