Real-time scheduling - Multiprocessor scheduling

Since the early 2000s, there has been a noticeable slowdown in the exponential growth (Moore's law) in computing speeds from one processor generation to the next. On personal computers, the increase in computing power is maintained by an increase in the number of computing cores on the same chip. This phenomenon can be explained by the miniaturization achieved by processor foundries on silicon transistor-based technology: for a given transistor size, accelerating computing speed has a quadratic cost in energy. Today, however, the size of a transistor is only a few atoms; it is therefore increasingly difficult to reduce, and the gains in computer frequency are less and less noticeable.

When ECUs are embedded in a system-controlled process (aircraft, automobile, train, artificial satellite, etc.), the trade-off between processing power and energy consumption is often in favor of limiting the computing frequencies of on-board processors when they are powered by battery, ambient energy or fuel. The reduction in processor power is then compensated for by multiplying the number of processors for the same overall computing power.

There are, of course, two ways of multiplying the number of computers: connect several computers via communication networks (these are known as distributed systems), or place several computer cores on the same chip (these are known as multiprocessor systems, or multicore systems when the computer cores are on the same chip). Of course, you can also combine the two and have communication networks linking multi-core systems.

On a complex critical system, such as a civil aircraft, the control system is made up of a distributed system of single-processor systems. Tomorrow's avionics architecture, and probably that of other types of complex embedded systems, must increase computing resources without increasing the complexity of the already highly complex communication network in these systems. As a result, tomorrow's mission-critical system architecture will consist of distributed multicore systems.

Critical embedded systems are generally subject to time constraints: each function must execute within a given time interval, despite the fact that computing resources are shared between several functions. These are referred to as real-time embedded systems, which must be not only functionally valid, but also time-valid.

The aim of this article is to present the problem of temporal validation of real-time applications on multicore processors. It can be seen as a sequel to