Analysis and Simulation of a Parallel Packet Switch for Satellite On-board Switching

In this paper we consider a packet switching system composed of X parallel switching planes operating independently and at a speed lower than the input lines. Arriving traffic is segmented into fixed length cells, then each cell is sent to one of the X planes, where it is switched to the correct output port and finally recombined with the other cells, coming from other planes, to reconstruct the original packet. This architecture, originally proposed by Iyer and McKeown [1], is referred to as a Parallel Packet Switch (PPS) and allows to design a switching fabric operating at a fraction of the line rate R. A PPS, with planes operating at rate r, must have at least k=R/r planes to avoid systematic packet losses. In [1] it was proved that a PPS can emulate the behavior of an Output Queue Switch (OQS) with the same buffering capabilities and the same number of ports. However, the centralized scheduling algorithm required to achieve this result can not be easily implemented in hardware, due to its complexity. In this paper we propose a Redundant Parallel Packet Switch (RePPS), i.e. a PPS with more than k planes, with a distributed scheduling algorithm, and multiplexing/demultiplexing stages without coordination buffers, which is a fair trade-off between performance and complexity. In particular we show that the minimum number n = X - k of redundant planes required to emulate an OQS with FIFO policy under any incoming traffic type is n = k2-2k+1. The distributed scheduling algorithm, which is the key component of the proposed switch, is presented and its performance, analyzed thru simulation, is discussed for a realistic fabric with a limited number of redundant planes. The results so far obtained suggest a possible application of this architecture for satellite on-board packet switches.