Communications - July 2012

an delay of 2.7ms, 75th percentile delay of 5ms, and is less than 90ms 95% of the total simulation time.

make drops less randomly distributed while CoDel gets randomness from the independence of drop intervals and packet arrivals.

Dropping the Right Packets

Although most network analysis today assumes connections with an unloaded 100-ms RTT, in practice RTTs vary. Unless a path includes a satellite link, RTTs in the one-second range are usually caused by bufferbloat, not the intrinsic path characteristics. At the consumer edge, few connections will have less than a 30ms RTT. Since CoDel’s interval is weakly related to RTT, we tested the effectiveness of a 100-ms setting over a wide range of likely RTTs and report on the range from 10ms–500ms.

Figure 8 shows results for a variety of traffic loads, sorted by RTT. Both CoDel and RED keep the median delay low, but CoDel has higher link utilizations and better drop-share fairness, showing that CoDel’s design is more effective at dropping the right packets. CoDel’s utilization is very close to that of Tail Drop (except for an RTT of 500ms) but with much less delay. CoDel’s performance metrics are not significantly different between 30ms and 200ms RTTs. Utilizations are slightly less and have a larger range as the RTT increases, because some traffic loads have difficulty keeping the larger pipe full. The 500ms RTT utilization shows more variation, the low end corresponding to single FTPs, which have difficulty keeping a very large pipe full.

Figure 9 compares the Jain fairness index for the source-drop shares of CoDel and RED for these runs. CoDel consistently outperforms RED for this metric. This seems to be, in part, because the changes to the original RED

Consumer edge

This scenario roughly emulates a consumer edge for two (symmetric) bandwidths: 512KB and 1.5MB. The load includes a two-way 64Kbps CBR ( VoIP-like), an infinite FTP as a download, Web browsing at a rate of two connections per second, and uploads of small FTPs—1MB with short idle periods ( 5 to 15 seconds, uniformly distributed) between. The table lists results, where “C” is CoDel and “T” is Tail Drop and each link direction is shown separately. Although CoDel never drops packets at a higher rate than Tail Drop, it keeps a much smaller queue and transfers similar amounts of data, offering encouragement for taming bufferbloat.

The experiments presented here
mainly consist of “forward” traffic
where all the data traffic is going in the
analyzed direction. Reverse traffic has
well-known issues of ack compression
or data pendulum, which tend to push
the delay up and the utilization down.
There are known mitigations to im-
prove the mixing of acks and data pack-
ets that will be performed in the home-
router implementation. Even the
unmitigated simulation experiments
showed acceptable performance.