Lossless static vs. dynamic reconfiguration of interconnection networks in parallel and distributed computer systems

Advances in the technology of dynamic reconfigurable hardware allow enhanced architectures of interconnection networks for large scale down to on-chip parallel and distributed systems. To archive a higher communication performance the topology of the interconnection network can be reconfigured at runtime to support the particular communication profile of the active applications. One important issue concerning the reconfiguration of the network is the prevention of packet losses. Altering the network topology leads to a new routing situation for the remaining packets. Two approaches are possible: emptying all buffers of the network before reconfiguration, called static reconfiguration, or rerouting of remaining packets after reconfiguration, referred to as dynamic reconfiguration. For the latter approach, a reconfiguration-aware routing scheme is required, which is presented in this paper. In addition, the static and dynamic approaches are discussed and compared concerning complexity and performance. It is shown that the increased efficiency of the dynamic approach becomes negligible, if current available technology is regarded.