Enabling Active Flow Manipulation In Silicon-based Network Forwarding Engines

Editor's Notes

• #3: Here is the outline of the talk. First I will identify several driving forces that led us in this direction of programmable networking Next, I review some basic functionality of a routing network element. Then I introduce our idea when we develop the AFM concept I will describe a framework for which AFM can be applied I will also describe several relevant examples using AFM and the platform Finally I conclude with a hint of what we go from here.
• #4: To me as a researcher (before sabbatical) : to be able to implement several of our new congestion control algorithms on a real router. For Nortel Networks (if I am not wrong): potential revenue generating direction by inventing and developing advanced technology/ By looking at the Internet from users’ perspective, service providers’ perspective and network providers’ perspective, we have identified several driving forces that steered us in this direction of research: Users want intelligent services Service providers want to differentiate their service by QoS, time to market, flexibility in managing their services Network Providers want to manage their services efficiently and economically. They want to sell, lease their resources at premium price. They want to sell bandwidth on-demand, etc.
• #5: Here is an example of introducing services on-demand. The environment allows users/service providers to dynamically, securely download new service to a network device. The service may be a new QoS mechanism, a congestion control algorithms, a new service differentiation scheme, a new monitoring scheme, etc
• #6: Above all we need programmability in network devices for introducing, enabling all kinds of intelligent services. What we need : a framework, a platform independent API.
• #7: Today’s Internet really consists of three types of networks: Residential/Enterprise Networks, Access Networks and Core Wide Area Networks. The characteristics of these are very different from each other in terms of bandwidth, scale, and quality of service. We need a flexible edge device that is intelligent enough to address this type of impedance mismatches. IP/Optical networks IP/Wireless networks
• #11: Software router. Everything is integrated and have to be processed by the CPU. Slow Separate control from forwarding Control for determining optimal routing paths and for management Forwarding is dedicated and hence is fast.
• #13: Configuring the system at start-up time. Occasionally Control and/or computation on every packet. Active packets or Capsule Control On Demand. This is what we are interested.
• #14: The interest is in flows and aggregated flows (classes of traffic) The interest is in the types of action that can be operated on the flow. Both Identification and Action must be performed in real-time
• #15: Database of what to be done based on SLA Database of possible filters of interests AFM defines a set of primitive flows and operation to obtain composite flows AFM defines a set of primitive actions Flow and Action can form an algebra in the most general sense. One can actually design machine with this algebra. The main interest is in identifying specific flows and applying actions to alter the behaviour in real-time.
• #20: The active flow priority change is a control-plane network service that applies AFM to alter the packet forwarding priorities of particular flows in real-time. The experiment procedure is as follows (see Figure 3b). At the beginning, the first TCP flow at a constant rate of 100Mbps is set up from Source 1 to the Destination through the Passport. The link bandwidth between the Passport and the Destination is 100Mbps at maximum. At time 1.3 seconds, the second TCP flow at the same rate from Source 2 is set up through the same link to the Destination. When they become stable, each claims nearly half of the link bandwidth (47Mbps). Then, the ORE on the Passport is instructed to activate the “active priority” service, which employs AFM to detect particular flows and increases the packet priority of the second flow at time 3.8 seconds. As expected, the receiving rate of the second flow (now with a high priority) increases and stabilizes at the desired bandwidth (70Mpbs) and the low-priority first one at a lower rate (24Mbps)
• #22: Regatta is another control-plane service that employs the Openet infrastructure and AFM for automated supervision [23]. Regatta stops, in a dynamic fashion, flows through routers when a node operation fails and leaves them to the Regatta (routing) supervision procedure. The Regatta supervision procedure handles the bypass with minimal service interruption to the user Traffic flows between the two end users go through nodes a, f, and e which is the shortest path. After the network operator installed Regatta at node f, Regatta begins to unfold itself outside of node f. In particular, it installs itself in the two adjacent nodes: a and e. Node f does not play any active role of control except in this bootstrap operation. It becomes the subject of the attentions by other reliable neighboring nodes. The Regatta-activated nodes a and e exchange periodic heartbeats between them to supervise the well being of suspected node f. The type and rate of the heartbeats are defined by the operator.