Comparison of overall link capacity utilization
for a variable number of ECN-capable TCP streams

Goal: to verify the dependency of the ECN gain on the TCP traffic profile injected. We assume that by increasing the number of TCP streams the overall burstiness of the aggregate stream increases.
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 (no Solaris hosts in use)

Test description:

1. Topology see Test A.
2. Parameters
• min threshold constant = 100 pack
• max threshold constant = 300 pack
• drop probability: [1/50, 1/5, 1/2]
• TCP stacks: ECN and SACK-capable and non-ECN and SACK-capable (Linux hosts only)
• test duration: 3 min
• Variable number of TCP streams (streaming mode): [4,8,32,64]
• minimum link capacity available: 50 Mbps.
3. Traffic profile
• TCP streams generated by iperf;
• no background streams, in the sense that all are ECN or non-ECN capable;

Summary:

• as known, an increase in the number of concurrent streams produces a considerable increase in link bandwidth utilization, both with and without ECN;
• for 4 and 8 concurrent streams, ECN seems to introduce a gain for mark probability values equal or larger to 0.1. In particular the most significant gain (1%) was obtained for a mark probability equal to 0.2.
• with 32 streams, ECN still introduces a penalty, which however is rather small and limited to a mark probability value equal to 0.1 or less. Instead with 64 streams, ECN always seems to introduce a little gain for all mark probability values.

Comments:

• The utilization gain achieved by increasing the number of parallel TCP streams (both with and without ECN) is confirmed, as it can be seen in Figure 1 and 2. This should produce an overall increase in burstiness of the aggregate TCP stream.

Fig 1 and 2: utilization increase produced by a larger number of concurrent TCP streams without and with ECN respectively.

·        As intuition suggests, when the number of parallel TCP streams is low, the probability of having an instantaneous queue size larger than the queue length (probability of having tail drop) is very low or null. As a consequence, the presence of a non-null marking probability with ECN for queue sizes in the range [min_threshold, max_threshold] causes a restrictive behavior which tends to reduce TCP performance even if the tail drop frequency is very very low. This is confirmed with 4 TCP parallel streams, but not by the curve with 8 streams.

Besides, we could be expected to see a similar behavior both for ECN e non ECN capable streams when the mark probability is very low. In fact, having a low mark probability means for RED mechanism a low packet loss when the average queue length is between min and max_threshold, while if ECN is also enabled means a low packet mark probability. However, there is a low probability of having congestion on the link, therefore the two mechanisms, only RED and RED+ECN, should have a similar behavior in terms of percentage of link utilization.

Fig 3: min/average/max of link utilization percentage with 4 and 8 concurrent streams (ECN or non-ECN capable)

• Figure 4 shows the percentage of link utilization for 32 and 64 parallel TCP streams. With 32 streams we can see that there is still a quiet penalty introduced by ECN for a mark probability lower than 0.1.

This penalty disappears by increasing the mark probability, and with 64 streams, ECN seems to give improvements even if not too significant.

.

Fig 4: min/average/max of link utilization percentage with 32 and 64 concurrent streams (ECN or non-ECN capable)