Heavy Hitter Detection on Multi-Pipeline Switches

Abstract

Recently, several applications have been designed and implemented to run entirely in the dataplane. However, most if not all the applications assume that network traffic traverses the same pipe, from ingress to egress inside the switch. While this seems to be a natural assumption, it does not hold for current programmable hardware that supports two to four pipes and network traffic is spread among the different pipes. As a consequence, several applications may not work properly in a multi-pipe architecture and need to be redesigned to fit into such architectural constraint. In this paper, we call the attention to this challenge and elaborate on an initial solution for counting heavy hitters (HH) in a multi-pipe hardware (MPHH). Our solution keeps the HH counter only in the egress pipeline while temporarily caching the hashes at the ingress pipeline. We then carry the hashes from ingress to egress by using data packets so that the HH are counted only in the egress pipeline. We present our design around this issue, the challenges observed so far and some initial results.