Traffic behaviour of interactive time-dependent thin streams on the modern Internet

The aim of the TimeIn project was to study the how the traffic patterns of application-limited flows behave in the current Internet with regard to the experienced latency. Studies have shown that thin streams, consisting of small packets with high inter-transmission times, often experience extreme latency because the design of the transport protocols and buffering schemes has been focused on supporting greedy streams, rather than thin. We wanted to identify the effects that cause thin-stream latency and design mechanisms in the protocols and/or the network to counter the detrimental effects. Such mechanisms would need to be incrementally deployable in the current Internet Architecture. As more and more of the traffic traversing the Internet are interactive or have timeliness requirements, it becomes important to ensure that all traffic patterns are treated in a way that no service must suffer extra delay due to flaws in protocol or system design. In the TimeIn project, we have analysed Internet traces and looked at each step in the journey of a data packet from it leaves the sending machine until it arrives at the receiver. We have identified some key problems that arise in the interaction between end-host protocols and network algorithms and designed mechanisms to reduce the latency created by these problems. We have designed several mechanisms that are in different stages of maturity, from being available in every new Linux installation through inclusion in the mainline Linux kernel to experimental new algorithms with the potential to change the way we perceive Internet Services. We have contributed to Internet Standards in order to ensure that our mechanisms are interoperable with industry worldwide and that the ideas behind them are sound and acceptable. The work in the TimeIn Project also led to an EU project, Reducing Internet Transport Latency (RITE), that served as a valuable collaborator throughout the project. Through the collaborators in RITE and TimeIn, we have created a network of the foremost International researchers on Internet Latency and helped bring Internet Latency back on the research and standardisation agenda.

Our current Internet systems was built to boost throughput, maximising efficiency for applications trying to move a lot of data. A large class of applications, however, does not produce large amounts of data. They send data when an action is performed or a short control message is to be conveyed. Such applications are very often time-dependent and/or interactive and very often produce what we call thin streams: Data flows with high interarrival time between packets and small packet sizes. Since the focus on network development has been on maximising throughput, the underlying network systems of the Internet is currently optimised for greedy streams: flows trying to send as much data as possible having large packet sizes and short interarrival times betwe en packets. In this process, latency has gradually been traded off for incremental increases in throughput. Thin streams are especially vulnerable to this tradeoff since: 1) the transmission patterns deviate from what the network is optimised for, 2) the y represent interactive and/or time-dependent applications whose quality of experience will quickly degrade, even with small increases in experienced latency. The TimeIn project aims to locate the sources of increased latency in thin-stream traffic and d evelop mechanisms that will reduce the experienced latency without having to change the basic infrastructure upon which the Internet is currently built.