Strategies for solution  

We see four possible strategies for dealing with these problems.

1. Do nothing, and wait for congestion to cause poor service which chases users to more expensive networks with better service. In the meantime, the nature of existing FIFO queuing implies that congestion will result in random allocation of delays and lost packets, meaning that many users will experience lower quality of service, even those who request only very small amounts of resources.

2. Implement sophisticated usage tracking, real-time pricing, or resource reservation schemes. We certainly expect that in the long run the community will redesign and reengineer networks to support mixed traffic profiles, including real-time continuous media traffic, with orders of magnitude more bandwidth capacity, complex resource reservation, accounting and pricing schemes. However these proposals may well require major redesign, preceded by considerable discussion [2] [6] [8] [7] [9]. In the interim we have a time window in which the Internet must accommodate traffic for which it was not designed.

3. Hope that service providers will be able to continue to upgrade the network ahead of demand, with no transference of resource consumption and upgrade costs to users. Never viable in the long-term, this strategy was tenable only during the period of plummeting prices for T1. That is, a 24-fold increase in capacity was affordable well before the possibility of a 24-fold increase in demand. A similar improvement in cost-performance for T3 leased circuits is unlikely to occur soon.

4. Provide an interim solution which can be implemented quickly to alleviate the impact of new patterns of traffic behavior. We propose such a solution here.

Many believe that Internet dependence on the altruism of the end systems has become unrealistic, and that protection needs to become a network-based mechanism implemented in routers, especially given the most recent TCP optimizations which support very large windows and thus the consumption of large fractions of total network bandwidth. Even the version of TCP [1] which supports these features took several years after the initial TCP deployment before widespread implementation. Similarly, it will take years to address the problems with new protocols and to broadly integrate appropriate solutions. Architecturally we believe we need a mechanism for the existing Internet that is consistent with the existing protocols and can be implemented with only minor changes to existing software.

In pursuit of incremental progress, we propose a strategy for the existing Internet, not in order to support new real-time multimedia applications, but rather to shield, albeit in limited way, the existing environment from applications and users whose behavior conflicts with the nature of resource sharing. In the future, the Internet needs its own self-defense mechanisms; our proposal is only an interim measure to address current and short term problems.

We targeted four goals in our proposal. First, we aim for an increased effectiveness in making use of available bandwidth. In particular we strive for multiple service classes on a single physical network. Second, we anticipate a need for the proposed scheme to be socially accepted, since one cannot easily mandate behavior on the Internet; one must rely on peer pressure or other incentives. Third, related to the second, we want a scheme which is equitable, at least in some loose sense. Fourth, it should be easy to implement and reward early local implementors even before the full system is in place throughout the global infrastructure.