Hashing on broken assumptions

Speakers:
Lorenzo Saino, Fastly

Widely used techniques such ECMP and LAG rely on hardware-based hashing performed on a packet five-tuple to uniformly spread traffic over different links while ensuring that packets belonging to the same flow take the same path to avoid packet reordering. These techniques operate under the assumption that the underlying hashing algorithm is correctly implemented and configured and uniformly spread traffic. But is it really the case? This talk gathers years of operational experience from the perspective of a CDN on hashing in the wild gone horribly wrong - from hardware limitations, to buggy software across both switches and middleboxes. Understanding these broken assumptions is critical when scaling load balancing functions at the network layer, as well as operating anycast networks.