Unified network architecture for scalable super-calculus systems

Abstract

A network architecture is used for the communication between elementary calculus units or nodes of a supercomputer to execute a super-calculus processing application, partitionable and scalable at the level of calculus power in the range of PetaFLOPS. The supercomputer comprises a plurality of modular structures, each of which comprises a plurality of elementary calculus units or nodes defined by node cards a backplane, a root card, and a node communication network of the switched fabric fat tree type; ii) a synchronization architecture comprising a plurality of distinct node communication networks, configured for the communication of specific synchronization information different from network to network and with different characteristics; iii) a re-configurable Programmable Network Processor that implements the nodes both of the n-toroidal network and those of the synchronization networks. The node communication networks of the n-toroidal type and of the switched fabric fat tree type can be used alternately or simultaneously for the transmission between the calculus nodes of the same type of data and information also in a configuration of system partitions in order to achieve the desired interconnection topology of the nodes.