Comparison of TCP performance with and without ECN

Comparison of TCP performance with and without ECN
(TCP stack SACK-capable)
Goal: to quantify the gain due to the use of ECN in terms of application throughput.
Equipment:

chain of intermediate ecn-capable routers (Cisco 7500 and C7200), IOS: 12.1(4)E (experimental version) - only first router is ecn-capable in this test
pool of Linux hosts, v. 2.4.9, SACK-capable
pool of Solaris 2.7 hosts, non SACK-capable

Test description:

Tolopology see Test 1.
Parameters
- Duration of each single test: [5, 30, 60, 180] sec
- mark probability: [1/50, 1/5, 1/2]
- type of reference stream: ECN-capable and non ECN-capable
Traffic profile
- 21 TCP streams (generated by iperf)
- 1 reference stream between two Linux hosts, we repeated tests with ECN enabled and not enaled; theoughput was estimated only for the reference
- 20 TCP streams (background), non ECN-capable, injected by the Solaris hosts

Summary:

troughput variability increases if ECN is enabled;
throughput gain seems to be larger in case of short-lived streams (5 sec)
throughput gain with ECN is appreciable only if the mark proability is non-negligible, for example larger than 0.1

Comments:

TEST 1: we compare the throughput of a single stream run for an interval of 5 sec, 1 min and 3 min, with and without ECN (see Figure 1, Figure 2 and Figure 3 respectively).

Fig 1: comparison of TCP performance with and without ECN for 5 sec streams

Fig 2: comparison of TCP performance with and without ECN for 1 min streams

Fig 3: comparison of TCP performance with and without ECN for 5 min streams In all cases it seems that for a small mark probability (0.02), performance without ECN is better than with ECN enabled. The understanding of this behavior requires further analysis. However, we can that for itnermediate mark probaiblity values (like 0.2), the perforamance with ECN is always better. For larger values the gain is less relevant.
These results are rather negative, in the sense that we couldn't demostrate the gain introduced by ECN in case of streaming TCP applications. This may be related to the small RTT experienced in this set-up. Future tests on WAN connections will be scheduled to evaluted the gain for long-distance streaming conenctions.

S.Alessandrini, T.Ferrari, March 22 2002