contributed articles

Doi: 10.1145/1506409.1506426

The passage of time is essential to ensuring
the repeatability and predictability of software
and networks in cyber-physical systems.
BY eDWaRD a. Lee
computing
needs time
moSt microproceSSorS Are

embedded in systems
that are not first-and-foremost computers. Rather,
these systems are cars, medical devices, instruments,
communication systems, industrial robots, toys,
and games. Key to them is that they interact with
physical processes through sensors and actuators.
However, they increasingly resemble general-purpose
computers, becoming networked and intelligent, often
at the cost of dependability.

Even general-purpose computers are increasingly asked to interact with physical processes. They integrate media (such as video and audio), and through their migration to handheld platforms and pervasive computing systems, sense physical dynamics and control physical devices. They don’t always do it well. The technological basis that engineers and computer scientists have chosen for general-purpose computing and networking does not support these applications well. Changes that ensure this support could improve them and enable many others.

The foundations of computing, rooted in Turing, Church, and von Neumann, are about the

transformation of data, not physical dynamics. Computer scientists must rethink the core abstractions if they truly want to integrate computing with physical processes. That’s why I focus here on a key aspect of physical processes— the passage of time—that is almost entirely absent in computing. This is not just about real-time systems, which accept the foundations and retrofit them with temporal properties. Although that technology has much to contribute to systems involving physical processes, it cannot solve the problem of computers functioning in the physical world alone because it is built on flawed technological foundations.

Many readers might object here. Computers are so fast that surely the passage of time in most physical processes is so slow it can be handled without special accommodation. But modern techniques (such as instruction scheduling, memory hierarchies, garbage collection, multitasking, and reusable component libraries that do not expose temporal properties in their interfaces) introduce enormous variability and unpredictability into computer-supported physical systems. These innovations are built on a key premise: that time is irrelevant to correctness and is at most a measure of quality. Faster is better, if you are willing to pay the price in terms of power consumption and hardware. By contrast, what these systems need is not faster computing but physical actions taken at the right time. Timeliness is a semantic property, not a quality factor.

But surely the “right time” is expecting too much, you might say. The physical world is neither precise nor reliable, so why should we demand such properties from computing systems? Instead, these systems must be robust and adaptive, performing reliably, despite being built out of unreliable components. While I agree that systems must be designed to be robust, we should not blithely discard the reliability we have. Electronics technology is astonishingly precise and reliable,