FEC Killed The Cut-Through Switch

Abstract

Latency penalty in Ethernet links beyond 10Gb/s is due to forward error correction (FEC) blocks. In the worst case a single-hop penalty approaches the latency of an entire cut-through switch. Latency jitter is also introduced, making latency prediction harder, with large peak to peak variance. These factors stretch the tail of latency distribution in Rack-scale systems and Data Centers, which in turn degrades performance of distributed applications. We analyse the underlying mechanisms, calculate lower bounds and propose a different approach that would reduce the penalty, allow control over latency and feedback for application level optimisation.