Paper Review: An Architecture for Differentiated Services

Reviewer: Kenneth Chin

This memo describes a evolutionary architecture for differentiated services. The architecture is entirely different from our current Internet in the sense that the later is a best-effort delivery network with very limited and simple quality-of-service (QoS) capability while the former is designed solely to accommodate different kinds of services. The differentiated services architecture is based on a simple model where traffic entering a network is classified and possibly conditioned at the boundaries of the network, and assigned to different behavior aggregates (BAs).

Differentiate services (DSs) rely heavily on routers to manipulate the incoming packets and outgoing packets. It is fairly easy to observe the overhead incurred in the router forwarding as the router has to classify every packet (for example, looking at the codepoint) and to condition every packet (for example, marking, re-marking, policing, shaping and dropping). What's complicating the situation is that every domain has it's service provisioning policy, each node has it's service level agreement (SLA), traffic conditioning agreement (TCA) and traffic profile (it is optional). The conditioning of a packet requires the conformance of all the above stipulations and this is not efficient enough. Since many management rules are carrying across the network, security issues (for example, authentication) arise. Furthermore, multicasting is another open question as to how can it be done efficiently without degrading the unicast traffic.

My view to this new architecture is that I really appreciate the idea of service differentiation, but it may not be the most proper way to achieve such a goal. It's failure is due to that fact that many rules are floating around and those rules are prone to attacks. I would suggest, if possible, that we standardize those rules to minimize the sharing of rules among routers and to reduce the comparison of rules within routers. This can be done by establishing a non-profit organization which is responsible for creating and registering the rules for different services. The administrators of the routers are then responsible for updating the rules in the routers if they want their routers to support or remove a particular service. In that sense, the routing performance would be improved. Secondly, the new architecture should focus on multicast rather than unicast in the design phase because unicast is a subset of multicast. Regarding the security issue, the ultimate solution would still be encryption of the packets.

I like the idea in the memo, but it doesn't give much contribution. Hence, it is a 2nd-grade paper.