Performance comparison of one short-lived TCP stream with and without ECN

 

• Equipment:
  • Pool of Linux hosts, v. 2.9.4: ECN and SACK capable
  • One IBM workstation, AIX v. 5.1: ECN and SACK capable (Amsterdam)
  • chain of intermediate ecn-capable routers (Cisco 7500 and C7200), IOS: 12.1(4)E (experimental version) -  in this test only the first  router is ECN capable

  Test description:

   

• Topology:
  topology consists in a chain of ecn-capable routers connecting a pool of TCP senders (on the left of the chain) and a pool of TCP receivers (on the right), like in this map
  Bandwidth bottleneck: 50 Mbps

   

• Parameters
  • min threshold constant =100 packets
  • max threshold constant = 300 packets
  • drop/mark probability : [1/50,1/5,1/2]
  • file size constant = 10KBytes
  • test duration: 120 sec

   

• Traffic profile
  • 20  TCP streams:

  • 1 referenced TCP streams between a Linux host in Italy and AIX workstation: ECN and SACK, generated by Mod_Netperf run in request/response mode;
  • 19 background TCP streams between Linux  hosts: ECN and SACK, generated by Iperf

Summary:

• The improvements, obtained with TCP ECN capable streams, are significant with every mark probability value (like on LAN).
• the number of transaction over time increases if the mark probability is greater.

Comments:

• Observing table 1, we can see that ECN gives a relevant contribute to the performance TCP stream , in terms of number of complete transaction over time.
• The confidence intervals show that the transactions grow with the increase of the mark probability.

Table 1: number of complete TCP transactions over time with one short-lived TCP stream, with and without ECN

Serena Alessandrini and Tiziana Ferrari, August 13 2002