From: Shobhit Raj Mathur (shobhit@cs.washington.edu)
Date: Tue Oct 26 2004 - 00:40:09 PDT
Explicit Allocation of Best-Effort Packet Delivery Service
==========================================================
This paper introduces the RIO (RED with In and Out) approach to provide QoS guarantees. This is referred to as the "allocated capacity"
framework. This framework can allocate bandwidth to different users in a controlled and predictable way in times of network congestion.
It allows the providers to charge different prices to users with different service requirements, which is not possible in the public
Internet today. In this mechanism the service providers can monitor their networks to detect whether their customers requirements are
not being met and if so by what extent. So this framework provides several advantages to the service providers (revenue, monitoring
etc.) and to the end users (QoS guarantees, differentiates service etc.).
Unlike the Fair Queueing approach, the "allocated capacity" framework is simple, scalable and flexible to provide different kinds of
service. There is no separation of traffic into different flows or queues. A customers profile maybe of the form "X is allowed to send
upto Y Mbps". As long as X sends less that Y Mbps, its packets are marked as 'in', but once it exceeds the rate the packets are marked
as 'out'. The packets of all users are aggregated into one queue tagged as either 'in' or 'out'. Under low levels of congestion packets
marked as 'out' will be discarded by the RIO algorithm. As the congestion increases a higher percentage of 'out' packets are dropped and
when the congestion crosses a particular threshold value 'in' packets are dropped as well. A packet drop is a feedback to the source to
reduce its throughput. There are two schemes described in the paper, a 'sender controlled scheme' and 'receiver controlled scheme' for
controlling the traffic.
Using the above framework, the service provider can provide the customer some assurance that it will be serviced according to his
profile and customers who pay higher can be given better guarantees. This is obviously a leap from the traditional Internet architecture
which is based on "best effort". Moreover, it does not demand much from the routers and can be deployed easily. I feel the important
issues of 'flow specification' and 'admission control' have not been touched upon. There is no mention of how the customer describes its
profile to the service provider and how the provider determines if it can satisfy a new customers request. Without admission control
this mechanism could be 'over subscribed' and result in poor performance. This defeats the very purpose of the approach. The framework
also assumes profile meters are located at the edges of the network and RIO is implemented in all the routers. I agree that any solution
which attempts to provide QoS guarantees would require some changes in the routers, so this framework does the best it can.
Overall, the framework can provide differential services to its customers while being simple and scalable. But it can only help
the class of 'tolerant and adaptive applications' and is far from providing a framework to solve the other QoS issues. The framework
allows providers to design widely different service and pricing models without having to build these models into all the switches and
routers by pushing the complexity to the edges. Hence I see it as providing QoS guarantees while satisfying the end-to-end argument to the
best possible extent.
This archive was generated by hypermail 2.1.6 : Tue Oct 26 2004 - 00:40:10 PDT