Multiprocessor system for realtime robotics applications

The paper presents the design and implementation of a multiprocessor, MRTA, having a multiple module architecture. Each module consists of a tightly-coupled set of processors connected to a shared resource by means of a multiport bus. A shared resource consists of memory, interrupt routing, and I/O for process control. The modules are connected through some of the member processors. With its modular architecture, MRTA is reconfigurable and expandable. Reconfiguration appears at two levels: the module interconnection topology and the type and interfacing of each shared resource. This architecture is suitable for multiple (parallel and sequential) and hierarchical formulations of realtime controllers, especially in the area of robotics. MRTA was built to enable performance evaluation of realtime controllers. The system software was developed to allow inputting of the control problem, i.e. a set of tightly-coupled equations, and the generation of the parallel algorithm and codes for mapping within MRTA modules. The system uses an efficient scheduling strategy to deal with large-scale problems such as robot dynamics. To demonstrate the performance of the system, a typical problem of robot control has been investigated. Performance evaluation was carried out to address the issues of scheduling, process monitoring and synchronization. The variation in task execution time, which causes loss of synchronization, was studied together with its implication on the processor finish times. It is shown, using an example, that the resulting loss in synchronization has a reduced effect on overall performance when the transfer time is moderate compared with the execution time.