Some recent research topics
===========================

(See papers for more details. The notes here are only a brief
outline.)


- Congestion control is being redesigned for new situations like
  cellular data networks. In cellular networks, the wireless link
  bandwidth and delay change dramatically. The traffic is alsi
  bursty. These constraints confuse TCP. Several researchers have
  proposed new transport protocols that are more suited to these
  scenarios.

- Congestion control for RDMA networks. RDMA (remote direct memory
  access) applications require very high bandwidth (several tens of
  Gbps) and very low latency (sub-microsecond). Regular TCP/IP stacks
  running in software cannot meet these requirements. Infiniband is a
  separate suite of networking products that implement the networking
  stack in hardware to meet stringent throughput and latency
  requirements. However, Infiniband equipment is not compatible with
  regular Ethernet-based LANs. There has been some interest on
  building networking where Infiniband applications run on regular IP
  networks, and congestion control in such networks is an interesting
  problem.


- Another problem: how to achieve fine-grained control over forwarding
  paths. One of the papers ("Fibbing") presents an interesting
  approach. Simple destination based routing gives you very little
  control on what paths the packets will take - they will always
  follow the shortest path to the destination. The traditional
  approach to changing the forwarding paths (to achieve load balancing
  or traffic engineering) was pinning flows to network paths using
  MPLS. SDN is another approach that uses a centralized controller to
  decide forwarding paths. However, centralization makes SDN networks
  less resilient to failures. The Fibbing paper proposes a midway
  approach - the system systematically introduces fake nodes and links
  into the routing protocol messages, in order to manipulate
  forwarding paths. The routing protocols and forwarding path
  computation is still distributed, only the computation of fake
  nodes/links is done in a centralized manner. The paper claims to
  achieve most of the benefits of forwarding path control that SDN
  gives, with fewer of the hassles.

- P4: a blueprint for OpenFlow 2.0. OpenFlow enabled hardware switches
  are only capable of matching on a fixed set of fields. On the other
  hand, P4 aims to make the fields to match flexible as well,
  achieving the true vision of SDNs.