Latency and Consistent Flow Migration: Relax for Lossless Updates

Consistency in network updates is a nascent research area, especially in the context of traffic engineering or Software Defined Networks. Various approaches have been proposed and implemented in the problem space of flow migration and congestion, primarily focusing on different flows not breaking the bandwidth capacities of the used links during updates. However, current network update techniques overlook the effect of flows congesting their own path during a network update due to latency on the links. Furthermore, while congestion will be resolved eventually after the network update, the buffers of the affected routers can be filled for a long time period, leading to the following paradox: a flow is moved to a path with less latency, but the latency stays the same! As flows are often migrated because of latency concerns, this is highly undesirable. We show that these effects occur already in a small topology in practice, causing packet loss due to overfull buffers. Furthermore, we prove that finding a lossless flow migration is NP-hard, already for a single (splittable) flow on directed acyclic graphs. Nonetheless, we can relax latency requirements to still obtain lossless flow migration. To this end, we show how to adapt current systems such as SWAN or Dionysus [SIGCOMM'13/'14], also developing our own polynomial time schedule algorithm, and discussing future consistent flow migration technique adaptations.