Equipment:

• Cat6500 hardware configuration
• pool of Solaris 2.7 workstations
• receiving machine of reference (non-background) traffic: dell03.cnaf.infn.it

Network topology:

• Testbed topology
• Policing/marking/scheduling logical topology

Router configuration:

• Policer configuration
• WRR configuration

Traffic profile:

• UDP background traffic: we enable a number of UDP streams injected by sunlab2 and sunlab3, so that this amount of background traffic is sufficient to completely congest the GigaEthernet interface to router 12000.
  Script
• Reference UDP stream: a precedence 1 stream (marked by Cat6500) is injected by sunlab4.
  Prec 1 rate: variable in rage [20, 80] Mbps
  Script

Test parameters:

• amount of reference traffic sourced by sunlab4
• policing CIR rate (and normal=maximum burst size)
• weight of queue 1 and 2

Summary:

• The assignment of the theoretical weight to a given queue provides very good protection from background traffic in case of congestion. However, in order to completely avoid packet loss on high precedence classes a much larger weight has to be assigned (what we call the real weight).
• The real weight for UDP traffic has to be several order of magnitude larger than the theoretical weight.

Comments:

1. Test A: The configuration of Weighted Round Robin (queue weight) is modified in order to determine the most appropriate weight in order to get a target rate, where the target rate corresponds to the amount of prec 1 traffic injected. When configuring a weight, we consider that this weight has to take into account overhead traffic produced by IP and UDP overhead, for this reason, the weight has to be larger than the amount of application data produced by the UDP source.

   In WRR weights are assigned to define relative bandwidths of the 2 main queues according to the following formula:

   
   bw_share(queue_i) = W(queue_i) / (W(queue_1)+...+W(queue_n))
   

   For the sake of simiplicity, we configure weight of queue 1 and 2 so that the sum of the two is 100, and the weight corresponds to the percentage of bandwidth assigned to that queue.

   Theoretical weight is computed according to the following rule:

   
   W_t(queue_i) = W(queue_i) / (W(queue_1)+...+W(queue_n)) and input_rate = W_t(queue_i) * 1000 Mbps\
   

   In other words the theoretical weight is the amount of bandwidth to be assigned to queue 2 so that bandwidth is equal to the amount of prec 1 traffic injected (including the overhead).

   In this test the theoretical weight allows the destination to get an amound of traffic rate which is almost equal to what configured at the source. Just a few packets are dropped by the scheduler, in fact the destination reports some packet loss in this case.

   We call real weight the minimum weight assigned to a given queue so that no drops are reported by the destination. We see from the following figures that in order to get an absolute isolation of prec 1 and prec 0 traffic (i.e. to avoid any loss on the prec 1 class) the real weight is much larger than the theoretical weight.

   Figure 1: relationship between theoretical weight and actual UDP rate as reported by the UDP source and by the UDP dstination. The two rates correspond.
   Figure 2: relationship between theoretical weight and real weight.

T.Ferrari and A.Mangiarotti, March 25 2002