JumboGen: Dynamic Jumbo Frame Generation for Network Performance Scalability

Network line speeds have increased at a significant rate. Unfortunately, network performance has not been able to keep pace with increases in line speed. This is due to the majority of packets being less than or equal to 100 bytes in addition to network routers not being able to scale well with the increased number of packets. In this paper we present our solution, JumboGen, an approach that will allow for a higher utilization of larger packet sizes on a domain-wise basis. Through simulations and experimentation, we show that the dynamic creation of jumbo packets decreases the number of packets processed by core routers and does not have an adverse impact on link utilization or fairness. The final result of JumboGen is a reduction in the number of packets seen by core routers which directly improves network scalability.

Packets that take the same path inside a domain can be aggregated together in some Jumbo packets to ease the life of core routers ... this is the main idea proposed by this paper. Indeed, one can ask the question of why to be limited to the 1500 bytes imposed by Ethernet (or other sizes imposed by other technologies), while IP and core routers can allow for larger packets? At the ingress of an Internet domain, packets going to the same exit point (i.e., following the same path inside the domain) are aggregated together up to some time limit – in order to not alter the end-to-end delay. Once at the egress of the domain, Jumbo packets are split and individual packets are forwarded in the classical way. The solution this way is incremental and each ISP can decide or not to do this aggregation as a function of how speed and loaded its core routers are.

This is a nice idea that sure will trigger discussions in both the research and engineering communities. The paper includes a correct evaluation of the solution and interesting discussion on how it will integrate with existing network solutions as QoS architectures, MPLS, ECN and Active Queue Management. Clearly and as all reviewers agree on, the way is open for further research in this direction as the evaluation of the solution with more traffic traces and topologies, and its adaptation to scenarios where the route to the next domain can go through multiple exist points or follow multiple paths inside the network. The issue of whether network operators can collaborate together to relay Jumbo packets is also to be studied.