Mediation Policies and the Tiered Internet

Now that we have reached a common understanding on how the Mediated Bandwidth Agreement can be implemented with existing peering and transit agreements, a more thorough examination of Mediation Policies is in order. Let's begin by taking a looking at the existing highway system as a congestion metaphor:

When networks congest, they don't operate too differently from what we see in the above diagram. Diamond lanes flow along nicely since traffic cops and regulations keep cheaters from congesting that portion of the highway. On the other hand, the other side of the roadway congests rapidly and heavily. The middle of the highway gets a bit better "best-effort" service than those truck lanes. Unfortunately, the highway planning commissions don't have an easy way to to implement a mechanism for shunting the best-effort traffic that congests the middle of the road off into the "truck lanes." If they did, the heavy truck and other high capacity, slow moving traffic could be throttled back to create a more pleasant traveling experience for us even if we didn't have a diamond lane sticker on our car.

Probably the most interesting aspects of this metaphor is in measuring the congestion the traffic causes. The highway planners could measure the average speed of the traffic in each of the three categories and plan capacity increases or policy changes for shunting that traffic. Unfortunately, we don't have adequate tools like this for IP networks today. There have been quite a few attempts at establishing objective measures for TCP/IP traffic and how it consumes associated capacity of the routers through the network. Commercially, so far, there has been little success in these sorts of ventures over the past ten years or so.

The primary challenges all of these ventures have faced has been the lack of standardization of that capacity. If the highway planners didn't have the ability measure average speed, there would be no way for them to easily determine when or how to expand roadway capacity. Similarly, the Internet has lacked standardized measures for bandwidth demand and capacity because of the wide variances in network equipment design, queuing managements systems and the like. Standards like DiffServ have suffered being too "coarse grained" as result because they could not adequately statistically bound the packet loss and delay.

It turns out the design of virtually all network equipment can be abstracted except for one relatively small component, the queuing system. There have been a wide variety of regiments to doing queuing to improve "fairness" for TCP flows traversing network bottlenecks. The most widely deployed model today is to do no queue management, usually known as "tail drop" (or sometimes "drop tail") followed by active queue management (AQM) techniques known as random early detect (RED) and weighted random early detect (WRED). It turns out all of these techniques can be readily gamed by new or different versions of TCP as I've written about previously. Theoretically, RED and WRED algorithms shouldn't easily be gamed by these new TCP systems. Practically, it is another matter. The AQM systems are extremely difficult for network operators to tune appropriately for their networks. This lack of tuning makes it a relatively straightforward process for the non-linear scientist types who are developing these new versions of TCP to game those bottlenecks as well.

Mediation Policies and Network Traffic

If we can create a standardized form of capacity, then we can extend those principles to incorporate objective measures for demand as well. The "diamond lanes" on this tiered service really aren't all that difficult to manage because the demand can be made to be very stable and easily quantified. The dynamicism and volatility of the best-effort traffic is not as conducive to creating easily standardized measures. Therefore, having a "new measure" for both that capacity and demand enables us to place a new control system in the network. We call the capacity side of this system a secure mediation controller (SMC) and the demand side, InterStream TCP mixed with traditional IP traffic. Effectively, this system will enable measurement of when that best-effort capacity is being under- or over-utilized. A very sophisticated control system uses that information to figure when and how to shunt that traffic over to the "truck lanes".

 

This "new measure" enables us finally to have a model in place which is quite similar to what we have on our telephony systems today. In several regions of the world, the equivalent of the public utilities commission would specify the maximum "call blocking probabilities" a telephone operator could have to offer service in their region. These probabilities were based on something called an "Erlang Model or Formula". These new measures of IP capacity and demand can effectively give us the same capability for the best-effort and premium Internet.

Jeff Turner

Comments

Post new comment