Study of the performance of an optical packet switch architecture with highly distributed control in a data center environment

We investigate the performance of an optical packet switch architecture with highly distributed control for interconnecting cluster switches in a data center environment. The optical packet switch under investigation can be scaled to a very large port count to interconnect a large number of cluster switches. Flow control is employed to regulate the packets transmission between the electronic buffers of the ingress and egress cluster switches. An important feature of the optical packet switch is that the switch can be re-configured in few nanoseconds regardless the port count. This is essential to minimize the end-to-end latency. Moreover, the limited contention resolution capability of the optical packet switch is compensated by the electronic buffers in the cluster switches. We numerically investigate the performance of an optical packet switch with 1024×1024 ports in terms of packet loss, throughput and latency in function of the electronic buffer capacity and number of retransmissions. Simulations results show that increasing the input buffer size allows for low packet loss at the expense of higher latency. On the contrary, limiting the number of retransmissions allows very low end-to-end latency but higher losses. For a system with 1024 in/out ports with a buffer size of 19 packets and a resend limit equal to 9, we obtain a packet loss lower than 10^-3 and latency around 800 ns considering an input load of 0.5.