Images account for a significant and growing fraction of Web downloads. The traditional approach to transporting images uses TCP, but this is overly restrictive for image data. Our analysis shows that the in-order delivery abstraction provided by a TCP-based approach prevents the receiver application from processing and rendering portions of an image when they actually arrive. Thus an image is rendered in bursts interspersed with long idle times rather than smoothly. This paper describes the design, implementation and evaluation of an image transport protocol (ITP) for image transmission over loss-prone congested or wireless networks. ITP improves user-perceived latency using application-level framing (ALF) and out-of-order application data unit (ADU) delivery, achieving significantly better interactive performance as measured by the evolution of peak signal-to-noise ratio (PSNR) with time at the receiver. ITP runs over UDP, incorporates receiver-driven selective reliability, uses a congestion manager (CM) to adapt to network congestion and is customizable for specific image formats (e.g., JPEG and JPEG2000). ITP enables a variety of new receiver post-processing algorithms such as error concealment that further improve the interactivity and responsiveness of reconstructed images. Performance experiments across a variety of loss conditions demonstrate the benefits of ITP in improving the interactivity of image downloads at the receiver.

Images account for a significant and growing fraction of Web downloads. The traditional approach to transporting images uses TCP, which provides a generic reliable, in-order byte-stream abstraction, but which is overly, restrictive for image data. We analyze the progression of image quality at the receiver with time and show that the in-order delivery abstraction provided by a TCP-based approach prevents the receiver application from processing and rendering portions of an image when they, actually, arrive. The end result is that an image is rendered in bursts interspersed with long idle times rather than smoothly. This paper describes the design, implementation, and evaluation of the Image Transport Protocol (ITP) for image transmission over loss-prone congested or wireless networks. ITP improves user-perceived latency using application level framing (ALF) and out-of-order pplication data unit (ADU) delivery achieving significantly better interactive performance as measured by the evolution of peak signal-to-noise ratio (PSNR) with time at the receiver ITP runs over UDP, incorporates receiver-driven selective reliability uses the congestion manager (CM) to adapt to network congestion, and is customizable for specific image formats (e.g., JPEG and JPEG2000). ITP enables a variety of new receiver post-processing algorithms such as error concealment that further improve the interactivity and responsiveness of reconstructed images. Performance experiments using our implementation across a variety of loss conditions demonstrate the benefits of ITP in improving the interactivity of image downloads at the receiver.

"Soft state" is an often cited yet vague concept in network protocol design in which two or more network entities intercommunicate in a loosely coupled, often anonymous fashion. Researchers often define this concept operationally (if at all) rather than analytically: a source of soft state transmits periodic "refresh messages" over a (lossy) communication channel to one or more receivers that maintain a copy of that state, which in turn "expires" if the periodic updates cease. Though a number of crucial Internet protocol building blocks are rooted in soft state-based designs --- e.g., RSVP refresh messages, PIM membership updates, various routing protocol updates, RTCP control messages, directory services like SAP, and so forth --- controversy is building as to whether the performance overhead of soft state refresh messages justify their qualitative benefit of enhanced system "robustness". We believe that this controversy has risen not from fundamental performance tradeoffs but rather from our lack of a comprehensive understanding of soft state. To better understand these tradeoffs, we propose herein a formal model for soft state communication based on a probabilistic delivery model with relaxed reliability. Using this model, we conduct queueing analysis and simulation to characterize the data consistency and performance tradeoffs under a range of workloads and network loss rates. We then extend our model with feedback and show, through simulation, that adding feedback dramatically improves data consistency (by up to 55%) without increasing network resource consumption. Our model not only provides a foundation for understanding soft state, but also induces a new fundamental transport protocol based on probabilistic delivery. Toward this end, we sketch our design of the "Soft State Transport Protocol" (SSTP), which enjoys the robustness of soft state while retaining the performance benefit of hard state protocols like TCP through its judicious use of feedback.

The Application Level Framing (ALF) protocol architecture [2] encourages application control over mechanisms that traditionally fall within the "transport layer", e.g., loss detection and recovery. Traditional ARQ-based reliable protocols for unicast (e.g., TCP) as well as multicast (e.g., Horus [30], RMTP [15], etc.) number data units sequentially to detect losses. Unfortunately, these transport-level sequence numbers do not permit receivers to flexibly tailor their reliability semantics. Achieving receiver-driven reliability is cumbersome in the existing "layered" architecture of the network protocol stack where the receiving application has no knowledge of how application-level objects map onto transport level sequence numbers. In this paper, we propose a new data naming scheme that exposes the structure of application data to the transport layer, thereby enhancing the expressibility of an applications' reliability and ordering semantics. We apply this data naming scheme to a reliab...

The development and deployment of a large-scale, wide-area multicast infrastructure in the Internet has enabled a new family of multi-party, collaborative applications. Several of these applications, such as multimedia slide shows, shared whiteboards, and large-scale multi-player games, require reliable multicast transport, yet the underlying multicast infrastructure provides only a best-effort delivery service. A difficult challenge in the design of efficient protocols that provide reliable service on top of the best-effort multicast service is to maintain acceptable performance as the protocol scales to very large session sizes distributed across the wide area. The Scalable, Reliable Multicast (SRM) protocol [6] is a receiver-driven scheme based on negative acknowledgments (NACKs) reliable multicast protocol that uses randomized timers to limit the amount of protocol overhead in the face of large multicast groups, but the behavior of SRM at extremely large scales is not well-understood.In this paper, we use analysis and simulation to investigate the scaling behavior of global loss recovery in SRM. We study the protocol's control-traffic overhead as a function of group size for various topologies and protocol parameters, on a set of simple, representative topologies --- the cone (a variant of a clique), the linear chain, and the binary tree. We find that this overhead, as a function of group size, depends strongly on the topology: for the cone, it is always linear; for the chain, it is between constant and logarithmic; and for the tree, it is between constant and linear.