A communication infrastructure for a million processor machine

The SpiNNaker machine is a massively parallel computing system, consisting of 1,000,000 cores. From one perspective, it has a place in Flynns' taxonomy: it is a straightforward MIMD machine. However, there is no interconnecting bus structure, and there is no attempt to maintain coherency between any of the memory banks. Inter-core communication is implemented by means of chip-to-chip packet transfer. Unlike conventional parallel machines, where packet-based communication is supported by a software layer, in SpiNNaker, the packet communication fabric is built in at the hardware level, and is the only mechanism whereby an arbitrary pair of cores can communicate. There is no unifying synchronisation system, and the packet delivery infrastructure is non-deterministic. In a number of application arenas - most notably neural simulation - this architecture is remarkably powerful, and supports techniques that can only be clumsily realised in conventional machines. In order to realise these advantages, a software layer - the routing system - is necessary to facilitate and choreograph the movement of the packets throughout the machine. The sheer size of the SpiNNaker machine makes conventional techniques difficult or impossible; the machine has to be largely self-organising. The routing tables that underpin the communication infrastructure can therefore be derived dynamically as a prologue processing phase. This paper describes the low-level software packet management system embodied in SpiNNaker. I