1. 
   Equipment:
   • Catalyst 6506, with GigaEthernet interfaces and IOS image: Version 12.1(8b)EX2
   • Transmitting host: Sun Solaris 2.7 (sunlab4)
   • Receiving host: Dell Linux host (dell03.cnaf.infn.it)

2. Test description:
   Policing is enabled on the input interface to the transmitting host.
   Input UDP traffic is constant.

3. Topology:
   the pools of end-systems are directly connected to the GigaEthernet interfaces of switch Cat65006 (testing does not involved additional devices and test traffic is just local). The structure of the topology is illustrated in the following figure:

4. Parameters:
   • CIR
   • Normal burst size
   • PIR

5. Traffic profile
   A single point-to-point UDP stream is generated according to the following command line:

   
   Tx: iperf -c 192.168.8.3 -t 200 -p 50006 -u -b 88.3 -i 10	// 90 Mbps, if overhead is included
   Rx: iperf -s -u -p 50006 -i 10
   	

   The application generates memory-to-memory data exchanges.

Summary:

1. TEST A (Normal Burst): normal burst size irrelevant to the applciation performance generating UDP traffic (throughput equals the CIR even for small buffer sizes like 64 Kby).
   • Test A: variable Normal burst size

     
     Command line:
     police 50000000 normal-burst-size maximum-burst-size conform-action transmit exceed-action
     

     Normal and maximum burst size in this test are set to the same value.
     In this test we see that for a normal burst size (=maximum burst size) in the range: [64 Kby, 2500 Kby] the achieved throughput as measured by the application is completely constant and equal to the bandwidth set by the CIR value. In particular, for a CIR of 50 Mbps, the corresponding performance measured is of 49.2 Mbps.

     This contradicts what seen with TCP traffic, since in that case, the relationship between normal burst size and committed information rate is linear.

   • Test B (normal burst size less than 64 Kby): in this test we try to focus on the following range of normal burst sizes: [2 Kby, 40 Kby]. We try to verify the exact minimum normal burst size allowed for a range of CIR values: [32 Kbps, 80 Mbps].

     For CIR values in the range [32 Kbps, 4 Mbps] the minimum configurable normal burst is only 1000 bytes, however, the minimum value which garantees an application throughput equal to the CIR is 2000 bytes.

     For CIR values larger than 4 Mbps, from 5 Mbps on, the minimum configurable normal burst size is the one accoding to the following table:

     CIR (Mbps)		normal burst size (kby)
     -------------------------------------------------
     5			2.5
     6			3
     10			5
     20			10
     30			15
     40			20
     60			30
     80			40
     

     If we try to set the normal burst size to a value slightly smaller than what indicated in the table (for example smaller by one unit), the CLI provides the following configuration error:

     31.25 Mby is the maximum configurable normal burst size.
     
     When adopting the configuration indicated in the table, the throughput
     reported by the application is optimum, in other words the achieved rate
     equal the CIR.
     
     
     
     
     
     Figure 1: minimum configurable normal burst size for different
     CIR values (UDP traffic)
     
     
     Test C (PIR): in this test we keep the maximum and normal burst size
     equal to 128 Kby and we vary the PIR rate, while the CIR is also constant
     and equal to 50 Mbps:
     
     
     police 50000000 128000 128000 pir pir[bps] 
     	conform-action transmit exceed-action transmit violate-action drop
     
     The input traffic profile is the same as the one specified for test A
     (the injected rate is 88.3 Mbps).
     
     Independently of the maximum burst size, we see that the achieved
     thrghouput as reported by the application is equal to the configured PIR - if
     we include the IP/UDP overhead which is not reported by the application.
     
     Figure 2, compares what achieved with UDP and TCP traffic. The absence of
     burstiness in case of UDP test traffic (that is CBR) justifies the good
     performance measured with UDP, which cannot be reproduced with TCP traffic
     due to the intrinsic burstiness produced by a TCP application running on a
     host with GigaEthernet interface.
     
     
     
     
     Figure 2: comparison of application performance as a function of the
     configured PIR with both TCP and UDP traffic.
     
     

   ---

   Tiziana Ferrari, March 25 2002