[e2e] QoS vs Bandwidth Overprovisioning

[Lijie Sheng]

Dear Ping Pan,

At 2001-04-25 14:07:00, you have written
>demir wrote:
>
>I doubt there is much disagreement from your above statement. One of the
>key issues is how to solve the problem (or from which direction should
>we deal with the problem). I'm an engineer, so sorry about my ignorant
>on architecture and religion. IHMO, much of the work in the past has
>been focused on QoS requirements and solutions from the end-user point
>of view. Granted that end-to-end has been very successful, but we have
>to realize that today's Internet is very different now (like it or not).
>It's a big business, so instead of designing something that can only
>satisfy end users, any deployable solution must first satisfy network
>and service providers. When you talk about optimizing network resources,
>the question is not whether you can come up with some beautiful
>equations, rather where do you want the resources to be optimized? At
>core? At edge? Would it work if the networks run different routing
>protocols with different policies? When you say scalability, what is not
>scaling in the first place? There was an argument on memory and hardware
>scaling, thanks to Moore's Law, we don't have such problem yet. 

	Hi, I have another question here, i.e., what resources are scarce in routers now?
Generally and abstractly, we can classify the resources in a router into there catagories, processing cycles, link bandwidths, and storage capacities. In the past, the conventional opinion was, all these resources are in short in equipment, so we have to manage all of them carefully; especially when considering the qos levels of different traffic, we have to distribute them carefully to fulfill the traffic with high qos level.
	However, now i guess that the storage capacities (memory) are fairly enough in today's routers. The only potential bottlenecks are processing cycles and link bandwidth (Here i ignore the fetching time of memory). A simple estimation:

	Considering an OC48 linecard, with speed 2.5Gbps. Assume the delay up-bound is 1s (extremely large i think). With little law we can estimate that the maximum starage will be

		2.5Gbps*1s=2.5Gb=312.5MB

	So 512MB memeory (DRAM) will fully satisfy the requirement. Actually, 256MB is fairly OK. Is 256MB DRAM too expensive or too large in space size for a 2.5G linecard?

	Thus, if memory is not the scare resource in router, when attempting to provide qos, we will not have to consider the partition issues of memory to different traffic. That is, manage the bandwidth and processing cyles with queue scheduling disciplines are fairly enough. Moreover, among qos requirements, only delay and throughput metrices need to be considered when designing the mechanisms and algorithms in router. 

	Am i right?  Maybe i missed a some key factors :)

>
>I read a paper from Dave Clark recently on rethinking e2e argument. It
>was very very good (sorry, I don't have the pointer.) 
>
>2 cents,
>
>- Ping


Best regards!

Lijie Sheng