One of the core design principles of the Internet is the idea that it is sufficient for networks to implement a minimal set of functionality (the IP protocol) to allow networks to interoperate and any two nodes to communicate. However, many recent network technologies (e.g. wireless, long-latency and frequently-disrupted links) have been difficult to integrate into the Internet for this reason. This project explores two core techniques that enable better support for unusual links: explicit path queries and end-to-end adaptation. Internet support for explicit path queries would allow end-points to learn key properties of the links used for their communication. The i-Path system from Waseda University and AIST in Japan provides the foundation for this path query mechanism. New end-to-end adaptation mechanisms would allow end-points to adapt their behavior to the capabilities or quirks of the communication path. In particular, the project is modifying two end-point systems to react to path query results: 1) TCP congestion control and retransmission and 2) multimedia streaming applications. This effort is expected to show how such systems should adapt to the path properties, such as latency, bandwidth, loss rate, router buffering and in-network compression support. The broader impact of this work comes from showing that explicit path querying is critical for simplifying the addition of new technologies to the Internet. By lowering the bar for supporting link new technologies, this research enables a number of systems that will have significant social impact, such as sensors for elderly homecare, and delay-tolerant links to underdeveloped regions.
The goal of this project has been to explicitly support diversity at the architectural level and, thus, change the nature of the Internet's narrow waist. By supporting diversity at the architectural level, the network can support new applications, node types and links in a simpler, more cost effective, and internally consistent manner. The mechanisms our work adds to the network architecture are (1) exposing to the end-points the "right" information about the properties of each link on the path and (2) allowing end-points to explicitly manage intermediary nodes on a network path. The work consists of two complementary, but related thrusts: Explicit path querying. Current Internet architectures make all links appear the same to higher layer protocols. Since this generic link requires reasonable ranges for minimum packet size, loss rates and bandwidth, some modern links fall outside the acceptable ranges. By introducing explicit path probing protocols, our design provides end-points with the specific information about the links their traffic traverses. End-to-End adaptation. Our work has shown how explicit path querying could enable a range of end-point based solutions that allow protocols and applications to adapt to path conditions. The focus of our efforts was on reliable transport protocols with an emphasis on timely loss recovery. In the first half of our work, we modified the i-Path protocol from our Japanese collaborators to play the role of the explicit path query protocol. Prior to our work, i-Path focused on queries about the performance properties of a path. Our work extended the i-Path design to accommodate richer queries about router features supported on the path – specifically queries to identify content-aware processing of packets. In the second half of our work, we developed a reliable transmission algorithm that works well over paths with routers that support content-aware processing. Historically, timely loss recovery has been accomplished using techniques such as forward-error-control (FEC). Unfortunately, FEC must be carefully tuned to the properties of the path, which has proven to be quite difficult. In content-aware networks, such as Redundancy-Elimination (RE) enabled networks and content-centric network architectures, the network caches content and effectively removes duplicate transfers of the same content. This provides a tremendous opportunity to use redundancy-based reliability schemes. However, redundancy must be introduced in the right way to ensure: (a) the network can eliminate it optimally to provide the desired efficiency and (b) the impact on other applications can be controlled. Based on this observation, we designed Redundant Packet Transmission (RPT) -- a scheme that intelligently sends multiple copies of the same packet. While sending duplicates copies of a packet instead of using FEC may seem counterintuitive and high-overhead, we found that RPT on content-aware networks can effectively support a variety of time-critical applications far better than existing approaches. When packets are not lost, the duplicate transmissions are compressed and add little overhead to the network. In contrast when the network is congested, the loss of a packet prevents the compression of a subsequent transmission. This ensures that the receiver still gets at least one decompressed copy of the packet. To illustrate the benefits of RPT concretely, we built a prototype for redundant real-time video streaming (RS) in a redundancy elimination network. Our evaluation of RS (presented in an USENIX NSDI 2012 conference paper) using a combination of real-world experiments, network measurements and simulations, shows that RS achieves far better video streaming than traditional FEC schemes. RS decreases the data loss rate by orders of magnitude more than FEC schemes applicable to live communications. As a result, it achieves better video quality than FEC for a given bandwidth budget, or uses up to 20% less bandwidth than FEC schemes to deliver the same video quality. Furthermore, it is simple and easy to use as it is much less sensitive to parameter selection than FEC. Finally, our congestion control design enables RS to coexist with existing traffic in a TCP-friendly fashion. Unfortunately, the RPT algorithm does not work well on networks without content-aware routers. Based on this observation, we have designed a transport protocol that uses FEC based reliability in existing networks and RPT in content-aware networks. The redesigned i-Path protocol provides the information needed to choose retransmission behavior. This part of the effort was critical to showing that a protocol like RPT could be deployed on the Internet. In summary, this project has made two significant contributions. First, it has shown how to support a diverse set of link and router functionality using a path probing protocol. Second, it has developed a novel reliability scheme for future content-aware networks. This has important implications for the future evolution of Internet functionality.
