Various experiments were tested at WAN-in-Lab, to develop FAST TCP and MaxNet at Netlab.
To enlarge diagrams, right click on the diagrams and select view image.

Alpha Tuning

Recent studies showed when heterogeneous congestion control protocols that react to different pricing signals share the same network, the resulting equilibrium may no longer be interpreted as a solution to the standard utility maximization problem. Although existence is still guaranteed, global uniqueness of equilibrium is lost. Hence it is hard to predict or even talk about steady state behavior, e.g. fairness, for such a system. Examples we considered included networks with TCP Reno and TCP FAST.

By tuning alpha parameter in FAST protocols, we can actually drive the system to have a global unique equilibrium which is also a solution of some weighted utility maximization problem. We developed an distributed algorithm to do that and experimentally study its performance. Preliminary results showed that it achieved very good fairness between FAST and Reno and also mainatained high throughput even when buffer is small.

Diagram of the single bottle-neck topology used for early Alpha-tuning experiments at WAN-in-Lab.

Loss Recovery

The loss recovery mechanisms in existing TCP variants were tested against the new loss recovery mechanism in TCP FAST. The experiment consisted of a WAN network with a wireless access link using an 802.11G wireless router. Loss was introduced by having another wireless router which generated traffic on the same radio channel as the access base station.

Diagram of single bottle-neck topology and wireless links used for Loss Recovery experiments.

MaxNet

MaxNet is a clean-slate re-design of Internet congestion control using theoretical analysis to guide the implementation. It is a framework based on explicit signaling of congestion information from routers to hosts. It could be a replacement of the source congestion control algorithm (eg Reno), the link AQM algorithm (eg RED) and an enhancement of the IP packet format.

MaxNet features a fully-distributed flow-control architechture, weighted Max-Min fairness, stability for networks of arbitrary topology, capacity and delay, fast convergence properties and quick starts for short flows. To achieve this, MaxNet uses an explicit congestion field in packet headers to communicate the level of the most congested link back to MaxNet senders. A MaxNet link AQM overwrites a packet's congestion value if its own congestion level is greater than the value already in the packet. MaxNet receivers intercept this signal and echo it back to the senders in ACK packets where it is then used to set transmission rate to an optimal value.

Some of our experiments used linux routers for the MaxNet AQM and modified linux TCP/IP stacks for the senders/receivers. This network topology allowed for several different experiments to be carried out testing MaxNet's fairness, link utilization and other properties.

Diagram of MaxNet testbed 2005.

©2005 California Institute of Technology - Networking Lab.