An adaptive, reconfigurable interconnect for computational clusters

This paper describes the principles of an original adaptive interconnect for a computational cluster. Torus topology (2D or 3D) is used as a basis but nodes are allowed to effectively migrate along the torus cycles. An optoelectronic scheme which makes such migrations possible with only local synchronisation is outlined. Between the instances of migration the interconnect behaves as a direct packet routing network which constantly monitors its traffic parameters. A decentralised predictive algorithm is applied periodically to decide whether the current topology is consistent with the predominant traffic flow and if it is not a reconfiguration to a better-matched topology occurs. We present simulation results that show that on some standard computational benchmarks a significant speedup is possible as a result of automatic matching between the effective topology of the application's message passing infrastructure and that of the interconnect.