Communication Aware Optimization of the Task Binding in Hardware/Software Reconfigurable Networks

Available networked embedded systems, e.g., in the field of automotive networks bind functionality statically onto electronic control units (ECUs). Thus, if a node fails, the functionality hosted by the control unit will be lost, and due to data dependencies, other functions on working nodes may not operate either. The same holds true for erroneous communication links which may lead to an isolation of a node or changing routes in point-to-point networks. However, in order to tolerate defects of computational nodes, it is necessary to replicate computational nodes and links such that a certain degree of redundancy is available. Obviously, introducing redundancy into embedded networks has an essential drawback concerning monetary costs, power consumption, size, weight, etc. Therefore, we will present an approach for tolerating permanent faults like node or link defects by separation of functionality from the physical hardware and rebinding tasks from defect nodes onto working nodes. When using reconfigurable devices such as FPGAs together with internal CPU cores, it will be possible to assign tasks implemented in either hardware or software, dynamically to the resources in the network. Besides these reconfigurable devices, computational nodes in a network contain dedicated analog hardware for driving sensors and actuators which leads to a certain heterogeneity in the network. This irregularity is typical for embedded systems that consist of specialized nodes for certain purposes. For such reconfigurable networks, we will show how to reduce the degree of redundancy on the one hand while increasing fault tolerance and flexibility on the other hand. Essential for achieving these objectives is a novel class of algorithms called online hardware/ software partitioning which determines a binding of tasks to available resources in heterogeneous networks. Binding tasks onto computational nodes has been investigated in many research fields. The offline approach towards hardware/software partitioning has been considered by many researchers [1][2][3]. For example, Blickle [1] synthesizes so-called Paretooptimal systems out of many design alternatives with the help of Evolutionary Algorithms. Such an approach helps a system designer for an unbiased decision making. Also, so-called load balancing algorithms have received a considerable interest and solve the problem of task binding with the objective of homogeneously distributing the task loads onto CPUs at runtime. Note that the term task is used in the same sense as ABSTRACT