Illustration <A NAME=illus> </A>



next up previous
Next: Special issues Up: Mitigatin the coming Internet Previous: Soft quotas on

Illustration  

 
Figure 1: Effective transmission success by precedence using a congested system (simulates precedence stratification in Table 1 on a T1 connection using a busy one-hour packet trace of traffic from UIUC connection into T3 NSFNET backbone during March 1993)  

To illustrate how precedence might work, we ran a simple simulation, using a one-hour trace of traffic at an inflow point of the T3 (45 Mbps) NSFNET backbone during March 1993. We divided the trace into one-second buckets, and assumed that the capacity each second was limited to 193,000 bytes, i.e., T1 (1.544 Mbps) capacity. If more than 193,000 bytes arrive in any second, the simulated router will drop some packets. Although this traffic was from the T3 network, we simulated its transmission through a link of T1 capacity to assess artificial congestion, allowing the illustration of the effect of our proposal. For this line capacity and our data set, the router dropped about 41% of the arriving bytes. We are thus simulating a heavily loaded system.

We compared the loss performance of two queuing rules: FIFO allocation, as routers now perform, and precedence allocation according to the priority assignment in Table 1. For the latter scheme we used the application information based on IP protocol and TCP/UDP port number of each packet in the trace, and simulated its transmission had it carried a precedence value according to Table 1. Figure 1 plots the success rate for all transmitted bytes for each precedence level. The dotted line indicates that under FIFO service (the present system), the router drops approximately the same fraction of packets from all precedence levels.gif The solid line indicates that in a precedence-based queueing system, in which a higher precedence packet always preempts a lower precedence packet, packets of highest precedence are subject to zero loss, while packets of the lowest precedence suffer extreme attrition.

This experiment represents just one example of a simulated environment, and other scenarios may be more reasonable depending on constraints of a particular environment. Examples include: (1) assuming packets rather than bytes were the bottleneck; (2) use a queueing rule that does not completely starve lower priority queues when serving higher priority queues; (3) apply an alternative stratification to that of Table 1.



next up previous
Next: Special issues Up: Mitigating the coming Internet Previous: Soft quotas on



k claffy
Fri Nov 25 20:51:38 PST 1994