max_threshold = min_threshold + deltawhere delta varies in the range: [40,60,100,200,300] The objective of the test was to identify the area in which the number of packets with CE bits set is relevant, i.e. the conditions under which the RED region [RFC 2884] dominates.
the percentage of CE packets is computed by taking into account the overall number of packets transmitted and by the number of packets merked as riported by the router.
In Figure 1, the number of packets with CE bit set is plotted. The number is expressed as a percentage of the overall number of packets transmitted (retranmitted packets included).
For low values of min threshold the percentage of CE marked packets seems to decrease. In the left area of the graph the RED algorithm is more reactive to potential congestion, especially when the difference between min and max threshold is small (40 bytes).
In this case, the number of CE packets may decrease as an effect of the prevailing effect of packet loss occuring when the average queue length is larger than the maximum threshold and of the fact that as a consequence the TCP rate is limited by the congestion control and avoidance algorithms.
When the min threshold is below a given threshold which depends on the delta in use,
there is no increase in the percentage of CE packets. This may be explained by the
fact that the aggregate TCP stream cannot produce an overall burstiness larger than
the threshold itself.
The larger is the difference between min_threshold and max_threshold, the larger
is the threshold at which a given curve becomes constant. A possible interpretation of
this is the following.
The drop probability configured in the router represents the mark probability
denomintator, in other words, the fraction of packets dropped when the average queue
depth is at the maximum threshold. This implies that the larger is the delta, the lower
is the probability of having a marked packet when the average queue length is equal to
a given value. In particular, as indicated in Figure 1.b, for each case in which
the average queue size is below point B, the difference in mark probability is equal
to Delta p. Delta p is larger for large differences between max and min_threshold.
However, for any instant in which the average queue size is larger than B, packets
are random dropped in case of curve 1, while they are still CE marked in case of
curve 2. However, if the frequency of such condition is small, the previous case
(average queue size less than B) prevails and the overall result is that with a small
difference between max and min_threshold the percentage of CE packets is larger,
as it seems to be suggested by Figure 1.
The delta between the two thresholds is constant and equal to 60 packets, while the packet drop probability is still constant and still equal to 1/10. The length of each test instance is 3 min.
Figure 1.b shows the throughput variation. While the curve is rather irregular, we can see that if the min_threshold is increased, then a small gain in throughput is achieved. This seems to confirm our interpretation of the fact that a small min_threshold tends to reduce the rate and burstiness of the ECN-capable streams, and consequently also the number of CE-marked packets, as shown in Figure 1.
Alternatively, the gain in throughput could be related to the number of CE-marked packets: when the number of CE-marked packets is small, similarly throughput is limited, while if the number of CE-marked packets increases, also performance improves.
This may be a proof of the advantage introduced by ECN.
As Figure 2 shows, mark probability is fundamental to increse the benefits of ECN (in terms of marked packets): the relation is according to a linear function.
